AXIS 1 · WHAT KIND OF THING IS IT?
Seven factory-shaped objects. Do not buy them as substitutes. Class is scope—not ownership. Availability says whether you can adopt it. Maturity says how far this exact surface has progressed. Neither changes where its work stops.
19 Operating factory A persistent system accepts bounded demand and repeatedly hands off a reviewable artifact through an evidenced production workflow.
27 Specialist factory A repeatable factory owns one bounded transformation family such as fleet migration, modernization, or incident-to-fix work.
22 Factory platform A control plane supplies durable orchestration, workers, policy, or isolation so an organization can assemble multiple factories.
14 Workflow kit An inspectable framework or workflow library defines repeatable agent steps but leaves fleet operation, enterprise containment, or release ownership outside.
8 Reference implementation / specification Public code or specifications teach one factory pattern but are not the evidenced production system they describe.
2 Assembled ecosystem Several separately owned tools form the candidate path; no single product owns the complete state machine or outcome.
3 Research prototype A paper, experiment, or time-bounded demonstration establishes a mechanism without a durable generally available operating surface.
◈ Enterprise internal Observed—not adoptable Strongest company-specific proof; you cannot adopt the factory.
Primary ↗ { } OSS / open Fork and operate You can inspect and fork it; you inherit operations and hardening.
Primary ↗ $ Managed product Buy access You buy the fleet and support; the control plane remains opaque.
Primary ↗ ⌁ Open reference Copy an architecture You inherit a design pattern—not maturity, security or support.
Primary ↗ ▦ Composable substrate Scope—not ownership You get machinery, not a complete factory or outcome authority.
Primary ↗ START HERE · THE AT-A-GLANCE FACTORY BENCH
Hold the work family still. Watch the systems diverge. “Software factory” is not one product category. Choose the nearest work family first; then read each system’s green edge and red cliff. Systems inside a family can still accept different inputs and stop at different artifacts—that stopping point is the comparison.
⇶ Fleet migration Can it repeat one governed change across many repositories? 3 ⟳ Legacy modernization Can it understand, transform, prove and cut over a brownfield system? 10 ⚒ Issue → PR Can it turn bounded product intent into a reviewable change? 23 ▦ Factory substrate Can we assemble and govern our own workers and workflows on it? 17 ⌁ Operations / domain authority Can it observe, govern or reconcile one live domain safely? 7 ◇ Specification / learning Can it turn judgment and failures into reusable factory memory? 2 ⚒ WORK FAMILY HELD CONSTANT Can it turn bounded product intent into a reviewable change?
23 relevant systems · public evidence orders the rows; it does not choose a winner
internal Stripe · Operating factory · internal only · production evidenced Stripe Minions Issue-to-PR maintainer
Primary evidence ↗ need same job
⚒ Code Can it produce reviewable software changes? ● leading →
✓ Prove Can a separate check reject it? ● strong →
⌁ Operate Can it close the loop in the live application? × absent →
⚖ Govern Can identity, policy, audit and recovery contain it? ● strong
× APP DAY 2 no
WHERE THIS FACTORY ACTUALLY STOPS Human-review-ready PR with CI evidence → NEXT ACCOUNTABLE OWNER Stripe code reviewer / owning engineer
✓ GREEN EDGE · CHOOSE IT FOR High-volume, well-scoped maintenance in a deeply instrumented monorepo with curated internal tools. × RED CLIFF · NEVER DELEGATE Portable adoption, broad language conclusions, release ownership, and application operations. ♙ HUMAN RESIDUE Humans open/take over work, review every PR, and retain merge and release authority.
One row · one executive fingerprint ✓ best use × fatal boundary ↻ Day‑2 boundary ♙ human residue Evidence is deliberately last. Documentation strength must not masquerade as suitability.
Factory / class / actual job ✓ Best use × Fatal boundary ↻ Application Day‑2 boundary ♙ Human residue Evidence Primary
internal operating Dropbox Nova Validated monorepo workbench internal only · production evidenced STOPS AT Validated branch change with validation results → Publication workflow / owning engineer Curated CI repair and migrations Private control plane; no production authority bounded Production crash alerts can seed test reproduction and a candidate fix routed to service teams; no live remediation/deploy authority. Humans own intent, architecture, quality review, release and final production judgment. ● E4 ↗ internal operating Duolingo AgentWorkflow Platform Durable reusable agent runtime internal only · production evidenced STOPS AT Typed workflow result with durable run artifacts → Calling workflow / registered agent owner Temporal retries, typed definitions and gated multi-surface agents Durable retries can repeat bad work; some targets cannot build in-worker bounded Agents fix CI, investigate PagerDuty/Sentry/Honeycomb evidence and support release managers, but writes require approval and humans retain incident and production authority. People approve write operations, review and merge generated PRs, supply domain judgment and handle production decisions; feedback is being converted into future evals. ● E4 ↗ internal operating Monzo Agent Chip Queued engineering and incident agent internal only · production evidenced STOPS AT Reviewable PR or incident-evidence bundle → Engineer / incident commander Permissioned workflows with bounded incident context Likely causes and draft fixes are not verified operational closure bounded Incident Mode gathers context, proposes causes, drafts fixes and has caught responder errors while humans coordinate; it does not own production mutation, rollout or recovery. Engineers and incident commanders judge causes, review/merge code, deploy and own customer impact; non-engineer contributions still enter normal review. ● E4 ↗ internal operating DoorDash Flux High-volume playbook factory internal only · production evidenced STOPS AT Playbook-declared run result: PR, finding or operations evidence → Invoking team / playbook owner Fast isolated workers and scoped reusable playbooks Task volume does not prove correct outcomes or safe release bounded Expanding from review into CI triage and on-call tasks; no unattended diagnosis-to-remediation-to-rollback loop is demonstrated. Teams author and tune playbooks, watch public run threads, evaluate output, review changes and retain merge, release, on-call and recovery authority. ● E4 ↗ internal operating WorkOS Horizon / TARS Event-driven internal code factory internal only · production evidenced STOPS AT Reviewable PR or structured investigation finding → Code reviewer / service or incident owner Context-rich triggered work in scoped cloud sandboxes Private recovery and no accepted-change or production-closure denominator bounded Can trigger from Sentry, slow-query and Datadog evidence, watch triage channels and draft incident-oriented work; humans retain diagnosis, production change, rollout and recovery authority. People approve decomposed issues, review/merge PRs, judge UAT, release and own incidents; self-improvement changes are human-reviewed PRs. ● E4 ↗ commercial operating Cursor Cloud Agents Durable cloud PR worker commercial product · generally available STOPS AT Draft PR with run and CI history → Human repository reviewer Isolated resumable cloud runs No multi-repository long runs no Cloud Doctor/security experiments and CI follow-up produce PRs; no generalized deploy-observe-remediate-rollback transaction. Normal terminal state is a draft PR; humans retain review, merge, deployment, and production authority. ● E4 ↗ internal specialist Meta SapFix + Getafix + Sapienz Historical crash-repair factory internal only · production evidenced STOPS AT Compiled, crash-replayed and test-validated patch candidate → Meta engineer reviewer Localize, generate, compile/test and rank Android crash patches 2018 narrow repair system; human approval mandatory and current availability unknown bounded It addressed pre-production crashes and static-analysis findings, including backlog cleanup. Engineers review every tested proposal, decide whether a nonroutine bug needs bespoke repair and authorize any production landing. ● E4 ↗ internal operating Razorpay Slash Context-rich multi-repo factory internal only · production evidenced STOPS AT Tested PR at the configured merge gate → Eligible automated gate or repository reviewer High-volume PRs, specialist review and empirical model routing Humanless merges lack a public defect and rollback denominator bounded Slash can inspect logs/Grafana, review support escalations, run operational reports and produce maintenance/CI PRs; no autonomous incident command, production rollback, SLO verification or direct production mutation is documented. Low-severity PRs can cross an agent-run merge gate without human comments. ● E4 ↗ internal operating Stripe Minions Issue-to-PR maintainer internal only · production evidenced STOPS AT Human-review-ready PR with CI evidence → Stripe code reviewer / owning engineer Scoped maintenance in Stripe’s curated monorepo Stops at a reviewed PR; humans merge, release and operate no Repository maintenance and CI repair, not deployed-service incident ownership. Humans open/take over work, review every PR, and retain merge and release authority. ● E4 ↗ internal operating Uber Minion Integrated background PR factory internal only · production evidenced STOPS AT Tested code change or PR → Owning engineer / code reviewer High-volume scoped work inside Uber's curated estate The broader SDLC portfolio is not one autonomous durable run no Recurring code maintenance, migrations, CI repair and performance campaigns are repository Day-2; no Minion-owned live detection-to-remediation-to-rollback loop is demonstrated. Engineers choose and structure work, supply legacy context, inspect architecture and semantics, approve/merge, plan rollout and own production incidents and business impact. ● E4 ↗ internal operating Vercel AI SDK Factory Multi-stage issue factory internal only · production evidenced STOPS AT Agent-reviewed PR with live-test evidence → AI SDK maintainer / reviewer TypeScript backlog automation No application operations no Backports and CI repair are repository Day‑2, not live service operations. Humans mark PRs ready and merge. ● E4 ↗ commercial operating Darkshift (formerly FactoryKit) Managed multi-repo issue-to-PR factory commercial product · generally available STOPS AT Evidence-backed PR for each changed repository → Human repository reviewer Isolated workers, brokered credentials and recorded browser QA Intentionally stops before merge, release and application Day 2 no A logged bug can start a fix run, but the product stops at a PR and explicitly cannot merge or push to the default branch. Humans define setup/check commands, review evidence and diffs, merge, release, handle migrations and own every production outcome. ● E4 ↗ commercial prototype Ona Software Factory Experimental lifecycle factory commercial product · prototype STOPS AT Merged or deployed revision with probe result, or escalation issue → Factory operator / product owner Build-to-monitor composition Only ten-day dogfood evidence bounded Post-merge smoke tests and a Sentry incident responder that opens issues; not autonomous incident command. Humans write specs, shape the factory, and review escalations. ◐ E3 ↗ internal operating Harvey Spectre Collaborative cloud-agent platform internal only · production evidenced STOPS AT Durable run record with reviewable artifact → Collaborating engineer / artifact reviewer Durable isolated runs with explicit permissions No public acceptance denominator or production authority bounded Can investigate incidents with approved telemetry and produce findings or code artifacts; no unattended production mutation, deployment or recovery authority is disclosed. People initiate and collaborate on runs, decide what failure means, review artifacts, approve follow-ups, merge, release and retain incident authority. ◐ E3 ↗ internal operating TRM AskNickiBot Factory Permissioned cross-repo factory internal only · production evidenced STOPS AT Reviewable cross-repository scaffold with launch evidence → Repository reviewers, then chain or on-call owner Shared scaffolding, review and launch evidence Humans retain architecture, launch and accountability bounded Produces on-call triage summaries, incident writeups and staging evidence, but humans decide customer flags, launch timing and production risk. Engineers own chain-specific logic and diffs; reviewers own correctness; chain owners and on-call leads sign launches; architecture and accountability decisions stay human. ◐ E3 ↗ internal operating Dexter Schwarm Reconciled background PR fleet internal only · production evidenced STOPS AT CI-green PR with live preview → Engineer / PR reviewer Small async tickets with CI and live previews Large ambiguous work collapses; no semantic oracle no Watches PR conflicts, CI failures, reviews and main-branch CI; it does not diagnose or operate the deployed application. Humans scope tickets, steer from the run view, inspect the live preview, review the PR, merge, release and own the application. ◐ E3 ↗ commercial platform Factory Droid runtime Embeddable coding-agent runtime commercial product · generally available STOPS AT Session result, structured output, branch change or caller-requested PR → Calling developer, workflow, reviewer or external orchestrator Resumable CLI, SDK and JSON-RPC sessions with tools and policy A session harness—not a durable factory, acceptance oracle or production controller bounded It can call operational tools when authorized, but Droid itself is an execution harness—not an incident truth oracle, rollout controller or production reconciler. The caller sets tools, policies and permissions and decides what to do with output; humans or external systems retain acceptance and production authority. ◐ E3 ↗ commercial operating GitHub Copilot Cloud Agent Hosted issue-to-PR agent commercial product · generally available STOPS AT Changed branch and optional single PR → Requesting developer / repository reviewer GitHub-native task governance Single-repository 59-minute ceiling no MCP can expose operational tools, but the product has no governed release, runtime-monitoring, remediation, or rollback closure. The agent cannot mark its PR ready, approve, or merge; requesting users cannot approve their own Copilot-authored PR. ◐ E3 ↗ open reference OpenAI Symphony Issue-to-worker scheduler open implementation · preview STOPS AT Workflow-defined accepted tracker or repository state with proof bundle → Adopter acceptance, review and merge system Inspectable minimal orchestration No durable claim or acceptance plane no Schedules repository work. Operators must harden sandbox, credentials, policy, acceptance, merge, release and recovery. ◐ E3 ↗ commercial operating Factory Software Factory Private-preview lifecycle control plane commercial product · preview STOPS AT Automation coverage/health state or workflow-specific run result → Configured downstream automation / software-team operator Manage and observe configured multi-stage Droid automations Dashboard state is not execution or outcome truth; no public acceptance, rework or safe Day‑2 denominator planned A private-preview incident workflow can investigate, draft an RCA or PR, and update a runbook; safe production effect closure is not established. People configure stages, credentials, tools and permissions; repository and production policy still decide review, merge, release and remediation. △ E2 ↗ commercial operating Netlify Agent Runners Web preview runner commercial product · generally available STOPS AT Build run with Deploy Preview → Developer / publisher Netlify-native web changes and deploy previews Web-delivery runner—not a general queue, verifier or Day‑2 operator bounded Deploy previews and web release context; not broad live-service remediation. Internal Builder can propose; Developers retain deployment/publication authority. △ E2 ↗ open workflow kit Mastra Factory Typed issue-to-PR factory open implementation · preview STOPS AT Reviewed PR with work item at terminal stage → Configured merge gate / repository authority Persistent typed handoffs Local sandbox is unsafe planned Monitoring/release agents are in the vision; default shipped path ends near issue completion. Persistent work-board transitions and configurable review/merge gates. △ E2 ↗ reference reference Vercel Eve Four-stage PR factory open implementation · preview STOPS AT Draft PR after analyst, implementer and reviewer stages → Human who marks ready / repository merge authority Forkable analyst → implementer → reviewer pattern Template lacks a qualified scheduler, governance and recovery plane no No application observability/RCA/remediation loop. Human makes the draft ready and merges. △ E2 ↗ AFTER THE FACTORY BENCH · EXPAND THE ECOSYSTEM Factories first. Components second. A reviewer, GitOps controller or SRE agent may be vital to a factory without being a software factory itself. Portfolio / build-buy-retire: Can it decide what to build, buy, retain or retire—and prove the real-world result? Demand / decide what matters: Can it turn noisy signals into one defensible priority—and stop work that fails?
Factory systems 95 Internal, OSS, managed and reference systems that organize software-producing work. Enabling systems 198 Reviewers, operators, identity/data controls, release gates and effect oracles. Whole ecosystem 293 Everything, when you are ready to design the complete technology organization. THE 20-SECOND ANSWER · 95 NONINTERCHANGEABLE SYSTEMS
Enterprise and OSS factories, with their edges exposed. Read great at, terrible at, then human. A factory earns only the narrow green job shown here. Its red edge and remaining operator outrank every average score or feature-list claim. The adjacent demand organ is separate: Similarity and record state do not prove dedupe, priority or outcome; Humans detect duplicates; votes are popularity, not priority. ServiceNow SPM Demand Management, Productboard + Pulse / Spark, OpenProject and Fider are decision-surface examples, not complete factories.
ACTUAL JOB FAMILIES 6 Compare within one first FACTORIES IN VIEW 95 Internal · OSS · managed NO APP DAY‑2 45 Stops before live operation BOUNDED / PLANNED DAY‑2 50 Narrow loop or stated intent DOCUMENTED APP DAY‑2 0 Still inspect its exact boundary
1 Hold the job constant Choose one work family. → 2 Read green, then red The cliff outranks the score. → 3 Inspect the residue Who still proves, merges, deploys and recovers?
⇶ Fleet migration Can it repeat one governed change across many repositories? 3 ⟳ Legacy modernization Can it understand, transform, prove and cut over a brownfield system? 10 ⚒ Issue → PR Can it turn bounded product intent into a reviewable change? 23 ▦ Factory substrate Can we assemble and govern our own workers and workflows on it? 17 ⌁ Operations / domain authority Can it observe, govern or reconcile one live domain safely? 7 ◇ Specification / learning Can it turn judgment and failures into reusable factory memory? 2
Find a named factory or boundary Show All system archetypes Enterprise internal OSS / open Managed commercial Open reference Factory substrate Rank by public evidence for overall evidence building proof application Day‑2 enterprise controls openness name Reset · 95/95
▦ Edge matrix Every system · same job, green edge and red cliff ⊞ Factory cards Scan by job family with expanded hover detail open commercial internal Side rail: application Day‑2
Factory passport What it actually is ✓ Great at × Terrible at / hard edge ♙ Human still required App Day‑2 Evidence fingerprint Primary
internal specialist Google large-scale LLM migration system internal only · production evidenced Fleet migration Nightly monorepo migration factory STOPS AT Compile- and regression-passing candidate change → Migration developer, then repository code owner GOOD Repeatable multi-language migration with indexed discovery and regression validation STOP Narrow ID-migration contract; humans still perform 25.55% of changes and approve every submission HUMAN Developers inspect leftovers, repair failed or hallucinated edits, update unsupported goldens, visually verify every change and send it to code owners for final acceptance. no ⚒ ● ✓ ● ⌁ × ⚖ ● E5 ↗ open specialist OpenRewrite OSS open implementation · generally available Legacy modernization OSS repository transformer STOPS AT Recipe-applied repository source diff plus data tables → Developer, repository CI and code reviewer GOOD Deterministic JVM/config semantic refactors in one repository STOP No governed fleet, runtime-equivalence proof, cutover or operations HUMAN Developers select/test recipes, run the build and CI, inspect diffs, review/merge, release and own behavior after deployment. bounded ⚒ ● ✓ ◐ ⌁ × ⚖ △ E5 ↗ open platform Agenhood open implementation · preview Specialized factory Fleet API/console → persistent agent container → task/workflow → snapshot/Git result STOPS AT Scheduled or ad hoc agent task with streamed events, workspace snapshot and optional Git change → Self-hosting operator, repository reviewer and target verifier GOOD A straightforward self-hosted fleet with persistent workspaces, pluggable Claude/Codex/OpenCode drivers, schedules, workflows, SSE, Git backup, MCP and filtered egress STOP Being multi-region enterprise control, application Day 2 or safe autonomous remediation HUMAN Self-hosting security review, credentials, skills/MCP, workspace roles, task policy, Git/release and target verification. bounded ⚒ ◐ ✓ ◐ ⌁ ◐ ⚖ △ E4 ↗ reference reference Agent-Native Runtime open implementation · preview Specialized factory YAML agent graph → mesh-mediated tools/delegation → HITL/egress policy → replayable audit STOPS AT Replayable agent-graph audit log and scenario result → Application owner, policy authority and independent deployment verifier GOOD Showing per-agent authority, data-context-aware egress review, bounded graph changes, human checkpoints and committed campaign logs as explicit runtime artifacts STOP Being a scalable production factory, distributed control plane, verifiable-credential system or independent business-effect oracle HUMAN Task/spec author, egress policy owner, HITL approver, tool/source owner, deployment authority and business verifier. bounded ⚒ △ ✓ ● ⌁ △ ⚖ △ E4 ↗ open platform AgenticOS open implementation · preview Specialized factory Versioned agent spec → governed runner → web/API/chat/automation → capability/MCP action → audit STOPS AT Versioned agent specification, governed run record and channel-delivered artifact → Agent owner, capability/identity authority and independent target verifier GOOD Centralizing versioning, exposure, budgets, approvals, capabilities, knowledge, MCP and audit across many user-facing channels STOP Cross-system source truth, target-specific authorization, causal RCA, application recovery or whole-IT effect closure HUMAN Agent owner, capability/identity owner and system-of-record owner approve exposures, side effects, budgets and recovery. bounded ⚒ ◐ ✓ ◐ ⌁ ◐ ⚖ ● E4 ↗ open workflow kit AOH open implementation · preview Specialized factory Org/team/role/skill pack → validator/compiler → Hermes/Claude Code/Codex workspace STOPS AT Validated Git-backed AOH pack and runtime-specific role workspace → Platform/operations owner and selected runtime authority GOOD Representing the IT department as portable roles, process skills, runtime requirements, model profiles and eval artifacts STOP a scheduler, worker fleet, permission system, deployment authority or Day 2 operator HUMAN Operations owners define roles, skills, model requirements, evals, host policy and target authority. bounded ⚒ △ ✓ ◐ ⌁ △ ⚖ ◐ E4 ↗ internal prototype Atlassian HULA / Rovo Dev internal only · production evidenced Specification / learning Human-guided coding pipeline STOPS AT Evaluation-passing PR candidate → Human developer / reviewer and merge authority GOOD Human-refined plans and evaluation STOP No unattended operations HUMAN Humans refine/approve plans and code; unit and functional evaluation follow. no ⚒ ● ✓ ● ⌁ × ⚖ ● E4 ↗ commercial platform Augment Cosmos commercial product · generally available Factory substrate Programmable SDLC factory STOPS AT Generated PR with verification report or incident finding → Human repository reviewer / incident commander GOOD Broad PR-through-dev-environment workflow STOP Its model review cannot independently prove behavior or safe effects HUMAN Humans decide merge, incident priority, remediation, production changes, and whether agent-generated evidence is sufficient. bounded ⚒ ● ✓ ● ⌁ ◐ ⚖ ● E4 ↗ commercial platform Cake Agents commercial product · preview Specialized factory Ticket/chat → private or shared Kubernetes session → OpenCode → tests/logs/artifacts → attributed commit or PR STOPS AT Session workspace, test/log/artifact bundle and attributed GitHub commit or pull request → Repository reviewer, independent CI/acceptance authority, then release and production operators GOOD Regulated engineering teams that want customer-cloud coding sessions, repeatable tools, central provider/repository controls, and live human collaboration around feature tickets STOP A complete software factory, autonomous RCA, production release owner, or whole-IT Day-2 operator HUMAN Humans own task intent and oracle, identity/repository/model/secret configuration, session steering, PR review, CI and semantic acceptance, merge, stage/production release, rollback, and incident command. no ⚒ ◐ ✓ △ ⌁ × ⚖ ◐ E4 ↗ open platform cat-factory open implementation · preview Specialized factory Board/tracker → durable pipeline → per-run container → review/test/acceptance → PR/CI/merge STOPS AT Verified PR and CI/acceptance evidence, or an explicitly escalated run report → Repository/release authority and independent production-effect verifier GOOD A broad, source-visible factory substrate with runtime-neutral domain packages, Cloudflare/Node/local facades, runner images, SDKs, spend/RBAC, model/agent registries and a PR-centered delivery loop STOP Proving autonomous whole-IT, production application Day 2, business/data/identity recovery or causal RCA simply because it opens a green PR HUMAN Requirements, risk, credentials, workspace/RBAC policy, merge/release, target identity/data acceptance, rollback and production effect verification. bounded ⚒ ● ✓ ● ⌁ ◐ ⚖ ◐ E4 ↗ open specialist Cloud SRE Agent open implementation · preview Specialized factory Log stream → threshold incident → triage/RCA/remediation → Go validator → PR/MR STOPS AT Sanitized incident analysis and validated Go patch delivered as file, PR or MR → Repository reviewer, release authority and service owner GOOD A clean ports-and-adapters separation with sanitized prompts, provider fallbacks, typed incident objects and a non-mutating PR delivery boundary STOP Live remediation, durable incident operations, arbitrary-language patch validation, target recovery or independent RCA HUMAN Humans review the analysis, patch, PR/MR, release and live recovery. no ⚒ ◐ ✓ ◐ ⌁ × ⚖ △ E4 ↗ commercial operating Cursor Cloud Agents commercial product · generally available Issue → PR Durable cloud PR worker STOPS AT Draft PR with run and CI history → Human repository reviewer GOOD Isolated resumable cloud runs STOP No multi-repository long runs HUMAN Normal terminal state is a draft PR; humans retain review, merge, deployment, and production authority. no ⚒ ● ✓ ◐ ⌁ × ⚖ ● E4 ↗ commercial operating Darkshift (formerly FactoryKit) commercial product · generally available Issue → PR Managed multi-repo issue-to-PR factory STOPS AT Evidence-backed PR for each changed repository → Human repository reviewer GOOD Isolated workers, brokered credentials and recorded browser QA STOP Intentionally stops before merge, release and application Day 2 HUMAN Humans define setup/check commands, review evidence and diffs, merge, release, handle migrations and own every production outcome. no ⚒ ● ✓ ◐ ⌁ × ⚖ ● E4 ↗ commercial specialist Datadog Bits AI commercial product · preview Operations / domain authority Signal → hypothesis tree → investigation → Bits Code PR / bounded remediation STOPS AT Investigation report, Bits Code PR or bounded remediation proposal → Service owner / code reviewer / separate rollout and effect authority GOOD The broadest documented observability-to-code surface: investigation, source PRs, security/cloud/test findings, adaptive detection and bounded Kubernetes remediation STOP Being treated as one coherent autonomous SRE authority HUMAN Humans curate telemetry/runbooks/memories, grant source permissions, review PRs and own rollout, compensation and business closure. bounded ⚒ △ ✓ ◐ ⌁ ● ⚖ ● E4 ↗ internal operating DoorDash Flux internal only · production evidenced Issue → PR High-volume playbook factory STOPS AT Playbook-declared run result: PR, finding or operations evidence → Invoking team / playbook owner GOOD Fast isolated workers and scoped reusable playbooks STOP Task volume does not prove correct outcomes or safe release HUMAN Teams author and tune playbooks, watch public run threads, evaluate output, review changes and retain merge, release, on-call and recovery authority. bounded ⚒ ● ✓ ◐ ⌁ △ ⚖ ● E4 ↗ internal operating Dropbox Nova internal only · production evidenced Issue → PR Validated monorepo workbench STOPS AT Validated branch change with validation results → Publication workflow / owning engineer GOOD Curated CI repair and migrations STOP Private control plane; no production authority HUMAN Humans own intent, architecture, quality review, release and final production judgment. bounded ⚒ ● ✓ ● ⌁ ◐ ⚖ ● E4 ↗ internal operating Duolingo AgentWorkflow Platform internal only · production evidenced Issue → PR Durable reusable agent runtime STOPS AT Typed workflow result with durable run artifacts → Calling workflow / registered agent owner GOOD Temporal retries, typed definitions and gated multi-surface agents STOP Durable retries can repeat bad work; some targets cannot build in-worker HUMAN People approve write operations, review and merge generated PRs, supply domain judgment and handle production decisions; feedback is being converted into future evals. bounded ⚒ ● ✓ ● ⌁ ◐ ⚖ ● E4 ↗ commercial specialist Dynatrace Davis / Intelligence commercial product · preview Operations / domain authority Grail/Smartscape problem → causal context → approved workflow action STOPS AT Correlated Davis problem, approved workflow result or runtime-context PR → Workflow approver / deployment, security or service owner GOOD The strongest declared causal/topology substrate in this cohort, with live problem lifecycles, dependency context, affected-user estimates and approval-gated workflows STOP Being treated as an independent application-effect oracle HUMAN Humans own instrumentation/topology quality, policy, approval, target scope, rollback/compensation and user/data/security probes. bounded ⚒ × ✓ ● ⌁ ● ⚖ ● E4 ↗ open workflow kit Fabro open implementation · preview Factory substrate Agent workflow graph STOPS AT Terminal run record with artifacts, checkpoints and judge result → Configured human node or adopter CI / release authority GOOD Auditable deterministic orchestration STOP No enterprise operations layer HUMAN Human nodes can be mandatory; adopter defines acceptance/release authority. no ⚒ ● ✓ ● ⌁ × ⚖ △ E4 ↗ open workflow kit FactoryOS open implementation · preview Specialized factory Intent → spec → plan → tasks → bounded agent work → verification → approval STOPS AT Repo-owned spec, task and verification bundle → Adopter-selected factory runtime / human approval authority GOOD Keeping product truth, execution truth, workflow commands, agent skills and approval gates in version-controlled repository artifacts STOP a hosted factory, worker scheduler, deployment system, identity plane, data plane or Day-2 operator HUMAN Runtime/integration choice, requirements, risk, credentials, tests, release, deployment and operations. no ⚒ △ ✓ △ ⌁ × ⚖ △ E4 ↗ open platform Gas City open implementation · preview Factory substrate Durable factory-builder SDK STOPS AT Gated merge-ready branch or PR with closed work record → Human merge / release owner GOOD Work ownership survives replaceable heterogeneous agents STOP Trusted packs, adopter-built gates and no application operations HUMAN Customers author Packs, trust boundaries, watchers and acceptance predicates; humans handle poison work, merge/release and production recovery. no ⚒ ◐ ✓ △ ⌁ × ⚖ △ E4 ↗ open workflow kit GitHub Agentic Workflows open implementation · preview Factory substrate Event-to-safe-output compiler STOPS AT Safe-output GitHub object applied by a separate job → Protected repository policy / human maintainer GOOD Brokered GitHub repository effects STOP Broker success is not semantic proof HUMAN Workflow owners constrain events, tools and effects; protected policy and reviewers retain semantic, merge, release and deploy authority. no ⚒ ◐ ✓ ● ⌁ × ⚖ ● E4 ↗ commercial platform GitLab Duo Agent Platform commercial product · generally available Factory substrate DevSecOps agent platform STOPS AT Draft MR with flow, CI and security results → Code owner / reviewer under project merge policy GOOD Integrated CI-security governance STOP No live-service remediation loop HUMAN Developer flows produce draft MRs and staged acceptance; humans retain approval and merge authority. no ⚒ ● ✓ ● ⌁ × ⚖ ● E4 ↗ open specialist k8s-mechanic open implementation · preview Specialized factory Cluster failure → stabilize/deduplicate → in-cluster OpenCode investigation → manifest validation → fix/investigation PR STOPS AT Fix or investigation PR with evidence, confidence and manifest-validation result → Human GitOps reviewer / merge authority GOOD A concrete, bounded failure-to-proposal loop with CRD state, severity, stabilization, read-only RBAC, secret/tool redaction, short-lived GitHub App tokens, dry-run and GitOps-only writes STOP General IT operations, causal certainty, business/user/data recovery, automatic production mutation or cross-platform model qualification HUMAN PR review/merge, GitOps reconciliation, application impact, secrets/RBAC, model policy and rollback/recovery. bounded ⚒ ◐ ✓ ◐ ⌁ ● ⚖ ◐ E4 ↗ open specialist Konveyor / MTA / Kai open implementation · generally available Legacy modernization Open migration assistant STOPS AT Grounded migration solution or IDE-applied source diff → Application developer GOOD Open rules-grounded assessment and candidate patches STOP Analyzer success does not prove build, behavior or cutover parity HUMAN Architects choose targets/rules; developers request and accept fixes, run builds/tests, deploy and own production. no ⚒ △ ✓ △ ⌁ × ⚖ ◐ E4 ↗ open specialist KubeRCA open implementation · preview Operations / domain authority Alert → evidence collection → RCA workspace → feedback/history STOPS AT Incident evidence packet, RCA workspace and operator-facing history → Incident commander / separate remediation and effect authority GOOD A durable incident context room that joins Kubernetes, Prometheus, Tempo and optional logs with Slack/dashboard history, feedback and similar incidents STOP Being treated as a causal oracle, autonomous repair controller or proof that a retrieved historical remedy applies to the current fault HUMAN Humans must sign and scope webhooks, validate evidence, adjudicate cause, authorize any mutation and own effect/recurrence closure. bounded ⚒ × ✓ ◐ ⌁ ◐ ⚖ △ E4 ↗ open platform Kubernetes MCP Guard open implementation · preview Specialized factory Kubernetes anomaly → bounded dry-run plan → OAuth/browser approval → digest-bound execution STOPS AT Digest-bound Kubernetes remediation plan, approval grant and executor receipt → Cluster owner, GitOps/release authority and independent workload verifier GOOD Making the approval, identity, plan digest, review digest and MCP mutation boundary concrete STOP General RCA, arbitrary Kubernetes operations, autonomous production changes or proving the workload recovered for users HUMAN Out-of-band OAuth-authenticated human approves the exact digest-bound plan; cluster and recovery owners remain accountable. bounded ⚒ ◐ ✓ ● ⌁ ◐ ⚖ ◐ E4 ↗ reference reference LangChain Open SWE open implementation · preview Factory substrate Open asynchronous coding-agent reference STOPS AT Draft PR from the persistent task sandbox → Adopter CI / human reviewer and merge authority GOOD Forkable intake, sandbox and draft-PR application STOP Validation is explicitly prompt-driven HUMAN Adopters configure allowed repositories and identity, add CI/review, inspect draft PRs, merge, release and operate applications. no ⚒ ● ✓ × ⌁ × ⚖ △ E4 ↗ open workflow kit Lvvphole Software Factory open implementation · preview Specialized factory Signal → triage/planning → bounded executor → tests/review/security/docs → verifier → draft PR STOPS AT Verifier-passing draft PR with run artifacts → Human repository reviewer / merge authority GOOD A compact, unusually explicit example of putting lifecycle, governance and pass/fail authority in code around a replaceable coding agent STOP Durable unattended operations, production deployment, SaaS/identity/data actions, semantic RCA or treating static diff checks as independent correctness proof HUMAN Signal placement, credentials, specification, waivers, merge, deployment and semantic acceptance remain human or external authority. no ⚒ △ ✓ ◐ ⌁ × ⚖ △ E4 ↗ internal specialist Meta SapFix + Getafix + Sapienz internal only · production evidenced Issue → PR Historical crash-repair factory STOPS AT Compiled, crash-replayed and test-validated patch candidate → Meta engineer reviewer GOOD Localize, generate, compile/test and rank Android crash patches STOP 2018 narrow repair system; human approval mandatory and current availability unknown HUMAN Engineers review every tested proposal, decide whether a nonroutine bug needs bespoke repair and authorize any production landing. bounded ⚒ ◐ ✓ ● ⌁ × ⚖ ◐ E4 ↗ commercial specialist Moderne Platform + CLI commercial product · generally available Legacy modernization Commercial fleet transformer STOPS AT Governed multi-repository campaign with patches, data tables, branches or PRs → Repository owners / migration campaign owner GOOD Governed multi-repository search, impact analysis and campaigns STOP No hidden-business-truth recovery, runtime proof, release or operations HUMAN Campaign owners select repositories and recipes; code owners inspect/merge PRs and platform teams release, observe and recover applications. bounded ⚒ ● ✓ ● ⌁ △ ⚖ ● E4 ↗ internal operating Monzo Agent Chip internal only · production evidenced Issue → PR Queued engineering and incident agent STOPS AT Reviewable PR or incident-evidence bundle → Engineer / incident commander GOOD Permissioned workflows with bounded incident context STOP Likely causes and draft fixes are not verified operational closure HUMAN Engineers and incident commanders judge causes, review/merge code, deploy and own customer impact; non-engineer contributions still enter normal review. bounded ⚒ ● ✓ ◐ ⌁ ◐ ⚖ ● E4 ↗ commercial specialist New Relic Autopilot commercial product · preview Operations / domain authority Alert → telemetry/knowledge investigation → reviewed remediation recommendation STOPS AT Telemetry-linked investigation and reviewed remediation recommendation → On-call engineer / workflow approval and target authority GOOD A read-mostly operational assistant with explicit reasoning display, telemetry investigation, knowledge/memory context and human-reviewed rollback recommendations STOP Being treated as a source-code factory or autonomous IT operator HUMAN Humans configure account scope and identity, review reasoning and rollback advice, approve workflows and verify target outcomes. bounded ⚒ × ✓ ◐ ⌁ ◐ ⚖ ● E4 ↗ open platform OpenGeni open implementation · preview Specialized factory API/webhook → durable session → managed sandbox or connected machine → approval/artifact STOPS AT Replayable session, approval decision and final artifact from a managed or connected-machine run → Calling product, target-system owner and independent effect verifier GOOD A serious open control plane for durable, replayable agent sessions, organization workspaces, governed credentials, connected machines, memory and artifacts STOP Owning target-system truth, causal RCA, safe production changes, identity recovery or whole-IT Day 2 HUMAN Deployment/access mode, sandbox profiles, credentials, approvals, target effect, recovery and business acceptance. bounded ⚒ ● ✓ ● ⌁ ◐ ⚖ ◐ E4 ↗ open platform OpenInspect / Open-Inspect open implementation · preview Specialized factory Signal or conversation → durable session → reproducible sandbox → coding harness → tests/artifacts → attributed draft PR STOPS AT Session artifact, tested branch or attributed draft PR with run evidence → Repository reviewer, merge/release authority and independent target verifier GOOD Providing a forkable session platform with multiple intake surfaces, per-session cloud development environments, provider choice, automations, multiplayer context, brokered secrets and reviewable repository output STOP Being treated as a complete organizational operating model, independent correctness oracle, multi-tenant SaaS, deployment authority or application Day-2 system merely because it can run agents and open PRs HUMAN Organization owners define repository scope, identity, network/secret policy, task routing, skills, semantic acceptance, PR review, merge/release and production recovery. bounded ⚒ ● ✓ ◐ ⌁ △ ⚖ ◐ E4 ↗ reference reference OurArk Genesis open implementation · prototype Specialized factory Trusted source body → inspect → descend/extend → inherited validation → independent lineage STOPS AT Validated descendant agent repository and lineage record → Human custodian / deployment and promotion authority GOOD Making an agent's behavior-defining repository, identity, private-state boundary and descent provenance explicit and human-custodied STOP Operating the descendant, autonomous IT, live remediation, production promotion or treating creation/recycle as a deployment system HUMAN Body selection, mission, permissions, private-state custody, deployment, promotion and semantic acceptance. no ⚒ △ ✓ ◐ ⌁ × ⚖ △ E4 ↗ internal specialist Ramp Inspect internal only · production evidenced Operations / domain authority Telemetry-driven bug repair STOPS AT Telemetry-linked reproduced fix proposal or PR → Owning engineer / code reviewer GOOD Telemetry-bound bug reproduction STOP No production authority HUMAN Humans define instrumentation, judge noisy thresholds, merge, deploy, and own production. bounded ⚒ ● ✓ ● ⌁ ● ⚖ ● E4 ↗ internal operating Razorpay Slash internal only · production evidenced Issue → PR Context-rich multi-repo factory STOPS AT Tested PR at the configured merge gate → Eligible automated gate or repository reviewer GOOD High-volume PRs, specialist review and empirical model routing STOP Humanless merges lack a public defect and rollback denominator HUMAN Low-severity PRs can cross an agent-run merge gate without human comments. Humans make repositories ready, encode tests/skills, handle exceptions and retain architecture, production risk and recovery ownership. bounded ⚒ ● ✓ ● ⌁ × ⚖ ● E4 ↗ open platform Runfusion Fusion open implementation · preview Specialized factory Intent → planner → workflow graph → plan/review/execute/review → worktree/PR/merge STOPS AT Workflow-governed repository change, PR or mission result with agent/model/event evidence → Repository reviewer, release authority and independent production verifier GOOD Operating a visible multi-project coding-agent fleet with PostgreSQL state, worktree isolation, selectable workflows, model lanes, missions, mailbox, plugins and mobile/desktop/web control surfaces STOP Owning application production operations, causal RCA, data/identity/security recovery or safe autonomous merge merely because a task passed internal gates HUMAN Product intent, workflow policy, provider credentials, repository authority, merge/destructive approvals, deployment and semantic acceptance. bounded ⚒ ● ✓ ◐ ⌁ ◐ ⚖ ◐ E4 ↗ commercial specialist Sentry Seer commercial product · generally available Operations / domain authority Error/performance issue → telemetry + code RCA → patch / PR STOPS AT Issue analysis, code-change proposal or GitHub pull request → Repository reviewer / CI / release authority GOOD A direct, narrow error-to-code path with explicit root-cause, solution, code-change, PR-iteration and handoff stages STOP Being treated as a complete brownfield factory, independent reproduction environment or proof that a generated PR repaired production HUMAN Humans scope repositories and telemetry, judge cause, review/merge code and own deployment, recovery and recurrence. bounded ⚒ △ ✓ ◐ ⌁ ◐ ⚖ ● E4 ↗ open workflow kit Simple Long Horizon Agent open implementation · preview Specialized factory Task → tools → plan/act/check loop → run trace → benchmark result STOPS AT Inspectable run trace and task result → Harness/evaluation operator or external target verifier GOOD A compact baseline for studying persistence, tool calls, verification, cost and model/harness differences without a large framework STOP Production-scale orchestration, enterprise identity, target effect, RCA, deployment, rollback or user support HUMAN Tool permissions, model/provider, task oracle, deployment and semantic acceptance. no ⚒ △ ✓ ◐ ⌁ × ⚖ × E4 ↗ open platform Smithers open implementation · preview Specialized factory Plain-English outcome → generated JSX workflow → agent/reviewer/eval → time-travelable run STOPS AT Time-travelable workflow run with reviewer/eval artifacts and proposed change → Repository reviewer, release authority and independent production verifier GOOD A highly inspectable coding-workflow workbench with live watch, rewind, fork, replay, reviewer sandboxes, benchmarks, evals and multiple coding harnesses STOP Being a durable enterprise scheduler, target-effect oracle or application operator HUMAN Workflow authoring, reviewer/eval design, repository review, merge/release and semantic acceptance. bounded ⚒ ● ✓ ● ⌁ △ ⚖ △ E4 ↗ reference reference Software Factory v2 open implementation · prototype Specialized factory Spec → locked interface → AC-derived tests → implementation → mechanical gate → cross-family review → model jury → integration → outcome verification STOPS AT Contract-bound tested module and outcome-verification bundle → Principal/specification owner and separate production release authority GOOD Contract-first handoffs, deterministic gates, failure routing back to contract revision and per-role/channel/model/prompt telemetry STOP Producing regulated production software, replacing brownfield systems or removing the final specification/outcome authority HUMAN Spec intake, work-item DAG creation, principal outcome review, regulated-domain exclusion, release and production operation. no ⚒ ◐ ✓ ● ⌁ × ⚖ △ E4 ↗ internal specialist Spotify Honk / Fleetshift internal only · production evidenced Fleet migration Fleet migration factory STOPS AT Per-repository migration PR or explicit no-op/failure → Repository owner / migration reviewer GOOD Standardized fleet migrations STOP Fails on irregular repositories HUMAN Humans define migrations/mappings, supply missing context, test ownership-sensitive behavior, review, and merge. no ⚒ ● ✓ ● ⌁ × ⚖ ● E4 ↗ internal operating Stripe Minions internal only · production evidenced Issue → PR Issue-to-PR maintainer STOPS AT Human-review-ready PR with CI evidence → Stripe code reviewer / owning engineer GOOD Scoped maintenance in Stripe’s curated monorepo STOP Stops at a reviewed PR; humans merge, release and operate HUMAN Humans open/take over work, review every PR, and retain merge and release authority. no ⚒ ● ✓ ● ⌁ × ⚖ ● E4 ↗ open workflow kit Super Simple Software Factory open implementation · preview Specialized factory Target repository + skill → deterministic Python graph → bounded agent nodes → specs/code/tests/review/docs STOPS AT Stamped workflow artifacts and target-repository change → Workflow adopter / repository reviewer GOOD A minimal composition pattern that makes the deterministic controller own sequencing while coding agents remain bounded workers inside a repository STOP Hosted factory operation, durable fleet scheduling, enterprise identity, live deployment, rollback, data workflows or broad language support claims HUMAN Installation, target commands, requirements, credentials, review, merge, deployment and operations. no ⚒ △ ✓ △ ⌁ × ⚖ △ E4 ↗ open platform SuperPlane open implementation · preview Factory substrate Agentic workflow control plane STOPS AT Completed versioned run with component outputs or release action → Downstream integration or human release owner GOOD Inspectable trigger, approval and integration graphs STOP Graph completion is not semantic or target-effect proof HUMAN People design graphs and predicates, approve protected paths, secure secrets/runners, interpret failures and own target recovery. bounded ⚒ ◐ ✓ △ ⌁ × ⚖ ◐ E4 ↗ internal operating Uber Minion internal only · production evidenced Issue → PR Integrated background PR factory STOPS AT Tested code change or PR → Owning engineer / code reviewer GOOD High-volume scoped work inside Uber's curated estate STOP The broader SDLC portfolio is not one autonomous durable run HUMAN Engineers choose and structure work, supply legacy context, inspect architecture and semantics, approve/merge, plan rollout and own production incidents and business impact. no ⚒ ● ✓ ● ⌁ × ⚖ ● E4 ↗ open platform useAgent open implementation · preview Specialized factory Web/Slack/REST/schedule → event-sourced thread → Linux workstation → approval/artifact delivery STOPS AT Event-sourced thread with sandbox artifact, approval receipt and delivered work product → Workspace administrator, repository/release authority and independent user-effect verifier GOOD An organization-facing coworker surface combining durable threads, isolated visible computers, typed gateway tools, Slack, knowledge/memory and native document artifacts STOP Being an enterprise-ready autonomous IT department HUMAN Workspace/connector policy, credentials, destructive approvals, artifact review, repository/release authority and user-effect verification. bounded ⚒ ● ✓ ● ⌁ ◐ ⚖ ◐ E4 ↗ internal operating Vercel AI SDK Factory internal only · production evidenced Issue → PR Multi-stage issue factory STOPS AT Agent-reviewed PR with live-test evidence → AI SDK maintainer / reviewer GOOD TypeScript backlog automation STOP No application operations HUMAN Humans mark PRs ready and merge. no ⚒ ● ✓ ● ⌁ × ⚖ ◐ E4 ↗ internal operating WorkOS Horizon / TARS internal only · production evidenced Issue → PR Event-driven internal code factory STOPS AT Reviewable PR or structured investigation finding → Code reviewer / service or incident owner GOOD Context-rich triggered work in scoped cloud sandboxes STOP Private recovery and no accepted-change or production-closure denominator HUMAN People approve decomposed issues, review/merge PRs, judge UAT, release and own incidents; self-improvement changes are human-reviewed PRs. bounded ⚒ ● ✓ ◐ ⌁ △ ⚖ ● E4 ↗ open specialist Xata Agent open implementation · prototype Specialized factory PostgreSQL signal → preset diagnostic/playbook → read-only finding → tuning proposal STOPS AT PostgreSQL investigation report with read-only diagnostic evidence and tuning proposal → Database owner and separate change/recovery authority GOOD Narrowing agent authority to source-native PostgreSQL diagnostics and explicitly avoiding destructive SQL STOP Broad IT operations, database mutation, migration, recovery, cross-system RCA or proving a user-facing fix HUMAN Database owner approves tuning, maintenance, migration and restore through separate authorities. no ⚒ △ ✓ ◐ ⌁ × ⚖ △ E4 ↗ open specialist Zelyo Operator open implementation · preview Specialized factory CRD policy/scan → security/anomaly/drift correlation → validated YAML/IaC → GitOps PR STOPS AT Correlated security finding with validated Kubernetes/IaC fix plan and GitOps PR → Security/platform owner, GitOps reviewer and independent target verifier GOOD Combining Kubernetes/cloud security scanners, compliance, drift, cost and GitOps remediation in a source-visible operator STOP Causal RCA, direct production mutation, complete security closure, business recovery or proof that broad scanner coverage is complete HUMAN Security/platform owners review findings, policies, PRs, merges, promotions and target recovery. bounded ⚒ ◐ ✓ ◐ ⌁ ◐ ⚖ ◐ E4 ↗ reference specialist Aegis open implementation · preview Specialized factory Metrics feed → observe/diagnose/propose → Pomerium gate → admin mutation → live verification → JSONL event log STOPS AT Gated remediation event log and live service verification result → On-call owner and independent recurrence verifier GOOD Testing denial feedback and re-planning: a tempting remediation is blocked, the reason becomes context, and a different action is attempted and checked against live telemetry STOP Generalizing a curated checkout service into autonomous SRE or assuming route-level authorization sees command-body risk HUMAN Fault injection, gate policy, fallback selection, scope of admin endpoints and independent recovery verifier. bounded ⚒ △ ✓ ◐ ⌁ △ ⚖ × E3 ↗ open workflow kit Agent OS open implementation · prototype Specialized factory Project chat → task graph → sandboxed coding agent → build/browser/vision verification → gated delivery STOPS AT Verified project task bundle with build/browser/recovery evidence and approval record → Project owner, deployment authority and independent production verifier GOOD The clearest new local end-to-end harness for turning model output into a verified project artifact through real builds, browser assertions, visual review and bounded recovery STOP Being a cloud factory fleet, enterprise IT operating plane or application Day 2 system HUMAN Task/plan approval, provider and connector scope, requirements, visual/semantic acceptance, deployment and recovery. bounded ⚒ ◐ ✓ ● ⌁ △ ⚖ △ E3 ↗ open specialist AgenticOps open implementation · preview Specialized factory Alert/request → HealthIssue → RCA → risk-tiered plan → approval → AWS executor → resolve STOPS AT Deduplicated HealthIssue with RCA, approved risk-tiered fix plan and execution/reporting receipt → AWS/service owner, incident commander and independent effect verifier GOOD Explicitly separating router, inventory, detector, RCA, plan, executor and reporting roles around AWS infrastructure STOP Independent RCA, cross-cloud production recovery, user/data/security closure and evidence of broad operational reliability HUMAN Humans own account scope, approval policy, L2/L3 actions, target recovery and semantic acceptance. bounded ⚒ ◐ ✓ ◐ ⌁ ● ⚖ ◐ E3 ↗ commercial specialist AWS Transform commercial product · generally available Legacy modernization Commercial modernization suite STOPS AT Job-specific output artifact plus worklog → AWS Transform Approver or Administrator GOOD Broad AWS-bound transformation STOP No whole-system equivalence proof HUMAN Humans authorize connectors, approve plans, supply dependencies, review code, test, confirm waves and own cutover. bounded ⚒ ● ✓ △ ⌁ △ ⚖ ● E3 ↗ commercial operating Cognition Devin commercial product · generally available Fleet migration Migration coding agent STOPS AT Completed session with branch, worklog and optional PR → Requesting engineer / repository reviewer GOOD Partitioned repetitive migrations STOP No autonomous production remediation HUMAN Humans resolve MFA/CAPTCHA and ambiguity, inspect changes, and retain merge/deploy/production decisions. no ⚒ ● ✓ ◐ ⌁ × ⚖ ● E3 ↗ internal platform Coinbase Mux internal only · production evidenced Factory substrate Agent-fleet concurrency controller STOPS AT Agent-authored branch or worktree changes → Initiating engineer / review and merge system GOOD High-concurrency agent orchestration STOP No reproducible verification HUMAN Engineers orchestrate agents and retain review/merge authority. no ⚒ ● ✓ ? ⌁ × ⚖ ? E3 ↗ open platform Config365 open implementation · preview Specialized factory Tenant template → Git/Gitea desired state → WhatIf/approval → act_runner/PowerShell → Microsoft Graph STOPS AT Approved WhatIf/apply run with Git-backed tenant configuration and target readback → Microsoft tenant administrator / independent user and recovery verifier GOOD Microsoft configuration control plane that packages tenant records, Git-backed baselines, runner execution, approval, backups and explicit update paths STOP Proving complete Microsoft workload coverage, safe tenant convergence, user/data/security recovery or production maturity from a five-commit public GA channel HUMAN Baseline/resource definition, Graph permissions, apply approval, runner/Gitea/SQLite operations, partial-effect reconciliation, recovery and user/business acceptance. bounded ⚒ ◐ ✓ ◐ ⌁ ◐ ⚖ ◐ E3 ↗ open specialist Cynative open implementation · preview Specialized factory Security question → authorized read-only fanout → sandbox analysis → cross-checked finding STOPS AT Read-only cross-system finding with source-linked evidence and provenance → Security/cloud owner and separate remediation authority GOOD A deliberate read-only-by-construction cross-system security and drift investigation surface STOP Remediation, target changes, continuous fleet scheduling, complete source inventory or business recovery HUMAN Security and cloud owners authorize read scopes, interpret findings and approve separate action planes. no ⚒ △ ✓ ● ⌁ × ⚖ ◐ E3 ↗ internal operating Dexter Schwarm internal only · production evidenced Issue → PR Reconciled background PR fleet STOPS AT CI-green PR with live preview → Engineer / PR reviewer GOOD Small async tickets with CI and live previews STOP Large ambiguous work collapses; no semantic oracle HUMAN Humans scope tickets, steer from the run view, inspect the live preview, review the PR, merge, release and own the application. no ⚒ ● ✓ ● ⌁ × ⚖ △ E3 ↗ open workflow kit DoorDash Agentic Orchestrator open implementation · preview Factory substrate Research → design → roadmap → implementation → antagonistic review → publish STOPS AT Published branch or PR with research and review artifacts → Human reviewer / merge and release owner GOOD Making long-running feature work explicit and resumable through deterministic phases, persistent artifacts, fresh contexts, bounded retries and independent-looking review roles STOP Being treated as Flux's public controller, a benchmark-proven autonomous engineer, a secure multi-tenant worker fleet, or a deployment and application-operations plane HUMAN People still establish requirements and repository authority, inspect trade-offs, resolve poison work, merge, deploy and own every production outcome. no ⚒ ◐ ✓ ◐ ⌁ × ⚖ △ E3 ↗ open workflow kit Elastic AI Software Factory open implementation · preview Factory substrate Repository workflow library STOPS AT Schema-constrained GitHub object with Actions evidence → Repository maintainer / reviewer GOOD Precision-first maintenance workflows STOP No durable workflow graph HUMAN Safe-output tools cannot merge or change repository administration; humans remain the review and merge gate. no ⚒ ◐ ✓ ● ⌁ × ⚖ △ E3 ↗ open platform EvoX Genesis open implementation · preview Specialized factory High-level objective → recursive responsibilities → transient agents/worktrees → validation → accepted repository world STOPS AT Accepted repository-world version with validation history → Human builder / semantic acceptance and promotion authority GOOD Persistent project state with disposable workers and recursive decomposition for extended software evolution rather than one long-lived model session STOP Enterprise application Day 2, identity/tenant/data authority, production deployment, target-side effects, business acceptance or assuming reported compiler/MESA demos generalize HUMAN Objective/constraints, credentials, policy, semantic acceptance, production promotion, data/security/identity and recovery authority. no ⚒ ◐ ✓ ◐ ⌁ × ⚖ △ E3 ↗ commercial platform Factory Droid runtime commercial product · generally available Issue → PR Embeddable coding-agent runtime STOPS AT Session result, structured output, branch change or caller-requested PR → Calling developer, workflow, reviewer or external orchestrator GOOD Resumable CLI, SDK and JSON-RPC sessions with tools and policy STOP A session harness—not a durable factory, acceptance oracle or production controller HUMAN The caller sets tools, policies and permissions and decides what to do with output; humans or external systems retain acceptance and production authority. bounded ⚒ ● ✓ △ ⌁ × ⚖ ● E3 ↗ commercial operating GitHub Copilot Cloud Agent commercial product · generally available Issue → PR Hosted issue-to-PR agent STOPS AT Changed branch and optional single PR → Requesting developer / repository reviewer GOOD GitHub-native task governance STOP Single-repository 59-minute ceiling HUMAN The agent cannot mark its PR ready, approve, or merge; requesting users cannot approve their own Copilot-authored PR. no ⚒ ● ✓ ● ⌁ × ⚖ ● E3 ↗ commercial specialist GitHub Copilot Modernization commercial product · generally available Legacy modernization Guided app modernization STOPS AT Assessment, plan or modified workspace with final report → Requesting developer / modernization lead GOOD Inspectable Java/.NET orchestration STOP No legacy behavior replay HUMAN Humans review assessment and plans, approve commands, edit deployment choices, review PRs and own release. no ⚒ ● ✓ ◐ ⌁ × ⚖ ● E3 ↗ commercial assembly Google Mainframe Modernization + Dual Run commercial product · generally available Legacy modernization Mainframe pipeline + Dual Run STOPS AT Migrated application slice with Dual Run parity report → Migration lead / cutover authority GOOD Separate online/batch parity plane STOP Not one autonomous factory HUMAN Humans validate rules, choose architecture/data model, review code, deploy both sides, configure comparisons and approve cutover. no ⚒ ● ✓ ● ⌁ × ⚖ ● E3 ↗ internal operating Harvey Spectre internal only · production evidenced Issue → PR Collaborative cloud-agent platform STOPS AT Durable run record with reviewable artifact → Collaborating engineer / artifact reviewer GOOD Durable isolated runs with explicit permissions STOP No public acceptance denominator or production authority HUMAN People initiate and collaborate on runs, decide what failure means, review artifacts, approve follow-ups, merge, release and retain incident authority. bounded ⚒ ● ✓ ◐ ⌁ △ ⚖ ● E3 ↗ commercial specialist IBM watsonx Code Assistant for Z commercial product · generally available Legacy modernization Mainframe transformation specialist STOPS AT Generated Java service or method with validation artifacts → Mainframe modernization developer / domain reviewer GOOD Method-level legacy/Java validation STOP No autonomous cutover or operations HUMAN Humans select services/classes/methods, approve actions, generate test data, judge parity and design the migration. no ⚒ ● ✓ ● ⌁ × ⚖ ● E3 ↗ open specialist Kubernetes AIOps Agent open implementation · preview Operations / domain authority Kubernetes evidence → multi-agent RCA → stable/canary decision → GitHub patch handoff STOPS AT Stable/canary analysis and GitHub issue or PR handoff → Rollout controller / code reviewer / independent verifier GOOD A concrete specimen for stable/canary evidence comparison, Kubernetes diagnostics and optional source patch/PR creation STOP Being trusted with rollout promotion or remediation because a model score, PR URL or nonempty response looks complete HUMAN Humans must force fail-closed rollout policy, validate root cause, build/test the patch, authorize deployment and independently probe user/data/security effects. bounded ⚒ × ✓ △ ⌁ △ ⚖ △ E3 ↗ commercial specialist Mechanical Orchard Imogen commercial product · production evidenced Legacy modernization Behavior-first modernization STOPS AT Modernized slice with old/new equivalence evidence → Modernization engineer / cutover authority GOOD Old/new production-data replay STOP Finite traces are not exhaustive HUMAN Humans review generated code, judge non-functional requirements, authorize cutover and retire each legacy slice. no ⚒ ● ✓ ● ⌁ × ⚖ ● E3 ↗ open workflow kit Microsoft Atlas open implementation · preview Specialized factory YAML/JSON workflow → Azure DevOps/AD/ARM REST calls → logs/rendered outputs STOPS AT Dry-run or executed Azure DevOps/AD/ARM workflow output with API responses and logs → Azure/DevOps administrator and independent target-state verifier GOOD Inspectable Azure-centric administration workflows with dry-run, retries, loops, secret redaction and no arbitrary shell execution STOP Being treated as Microsoft's autonomous IT department, a backup/restore system, an LLM factory, or an independent target-state/business-effect oracle HUMAN Workflow authoring, API scopes/service connections, dry-run approval, target readback, rollback/compensation and recovery. bounded ⚒ ◐ ✓ ◐ ⌁ ◐ ⚖ ◐ E3 ↗ open assembly NTM / Agentic Coding Flywheel open implementation · preview Factory substrate Multi-agent factory cockpit STOPS AT Operator-accepted revision or batch with verification provenance → Expert operator / separate merge and release authority GOOD Expert-operated heterogeneous swarm coordination STOP Advisory locks and cross-tool recovery prevent unattended authority HUMAN Operator owns decomposition, dispatch, acceptance, batch verification, and risky actions. bounded ⚒ ● ✓ ● ⌁ △ ⚖ △ E3 ↗ commercial prototype Ona Software Factory commercial product · prototype Issue → PR Experimental lifecycle factory STOPS AT Merged or deployed revision with probe result, or escalation issue → Factory operator / product owner GOOD Build-to-monitor composition STOP Only ten-day dogfood evidence HUMAN Humans write specs, shape the factory, and review escalations. bounded ⚒ ● ✓ ● ⌁ ◐ ⚖ ◐ E3 ↗ open specialist oncall-agent open implementation · preview Specialized factory Momento incident → dedupe/queue → Teleport auth → Bedrock investigation → ranked hypothesis → optional GitHub PR → Slack report STOPS AT Incident evidence bundle and optional remediation PR → Service owner, code reviewer and independent effect verifier GOOD Making incident identity, short-lived AWS access, tool-scoped investigation, structured state transitions and PR-only remediation explicit STOP Being treated as a causal RCA oracle or direct production repair system HUMAN IAM/Teleport scope, incident policy, remediation enablement, PR review, merge/release and independent target verifier. bounded ⚒ ◐ ✓ ◐ ⌁ ◐ ⚖ ◐ E3 ↗ open reference OpenAI Symphony open implementation · preview Issue → PR Issue-to-worker scheduler STOPS AT Workflow-defined accepted tracker or repository state with proof bundle → Adopter acceptance, review and merge system GOOD Inspectable minimal orchestration STOP No durable claim or acceptance plane HUMAN Operators must harden sandbox, credentials, policy, acceptance, merge, release and recovery. no ⚒ ◐ ✓ △ ⌁ × ⚖ × E3 ↗ open platform Owain Lewis Factory open implementation · preview Factory substrate Agent-fleet scheduler STOPS AT Published worker attempt with branch or PR changes → Repository reviewer / merge authority GOOD Durable fleet scheduling STOP No semantic correctness proof HUMAN Operators provision workers and credentials; humans review and merge the resulting PR and retain release/production authority. no ⚒ ◐ ✓ △ ⌁ × ⚖ × E3 ↗ open platform Polos open implementation · preview Specialized factory Agent definition → durable workflow → sandbox/tool call → approval/trace/artifact STOPS AT Durable workflow result with tool/approval trace and sandbox artifact → Workflow owner, repository reviewer and separate target authority GOOD A compact Python/TypeScript agent runtime that packages sandboxing, durable retries, approvals, cron/webhooks/events, Slack and OTel around user-authored agents STOP Being an independent correctness or RCA authority HUMAN Agent/tool definitions, allowlists, approval rules, sandbox images, credentials, target effect and recovery. bounded ⚒ ◐ ✓ ◐ ⌁ ◐ ⚖ △ E3 ↗ commercial operating Replit Agent commercial product · generally available Specialized factory Prompt/spec → cloud workspace → app artifact / deployable project STOPS AT Task result or project artifact with behavioral evidence → Project owner / reviewer / separate deployment and operations authority GOOD Fast browser-native application prototyping with an integrated workspace, preview and deployment path STOP Being treated as a governed enterprise software factory, causal incident investigator or independent production-recovery authority HUMAN Humans specify requirements, review code and permissions, configure integrations, approve deployment and own production support. no ⚒ ● ✓ △ ⌁ × ⚖ ◐ E3 ↗ reference reference StrongDM Factory / Attractor specification only · preview Specification / learning Factory specification framework STOPS AT Versioned NLSpec graph or specification document → Runtime implementer / factory operator GOOD Precise graph-spec vocabulary STOP Executable public runtime absent HUMAN Human-gate nodes are explicit in the public specification; the private no-human-review claim cannot be independently audited. no ⚒ △ ✓ ● ⌁ × ⚖ × E3 ↗ open platform Sympozium open implementation · preview Factory substrate Kubernetes agent operating plane STOPS AT AgentRun result, trace and delegated artifact from a policy-scoped Kubernetes Job/Pod → Calling workflow, evaluator or human operator; separate target authority GOOD Ephemeral runs, delegation and policy/sandbox hooks STOP AgentRun success is not RCA or target recovery HUMAN Platform owners must define model/tool/namespace policy, sandbox runtime, service accounts, target authority and the acceptance contract. no ⚒ × ✓ × ⌁ × ⚖ △ E3 ↗ internal operating TRM AskNickiBot Factory internal only · production evidenced Issue → PR Permissioned cross-repo factory STOPS AT Reviewable cross-repository scaffold with launch evidence → Repository reviewers, then chain or on-call owner GOOD Shared scaffolding, review and launch evidence STOP Humans retain architecture, launch and accountability HUMAN Engineers own chain-specific logic and diffs; reviewers own correctness; chain owners and on-call leads sign launches; architecture and accountability decisions stay human. bounded ⚒ ● ✓ ◐ ⌁ △ ⚖ ● E3 ↗ commercial specialist vFunction commercial product · generally available Legacy modernization Runtime-informed decomposition STOPS AT Generated Java service-code baseline → Application developer / extraction team GOOD Java/.NET architecture discovery STOP Extraction output is only a baseline HUMAN Humans exercise flows, choose boundaries, configure targets, refine generated baselines, test, deploy and adjust decomposition. bounded ⚒ ◐ ✓ ◐ ⌁ △ ⚖ ● E3 ↗ commercial platform Warp Factories / Oz commercial product · preview Factory substrate Factory-as-code control plane STOPS AT Workflow-defined result; reference flows emit verified PRs → Specification or code reviewer / merge authority GOOD Multi-harness cloud workers STOP Managed control plane opaque HUMAN Versioned checkpoints, session steering/handoff, human spec review and merge in the reference flow. planned ⚒ ● ✓ ● ⌁ △ ⚖ ● E3 ↗ open reference ZeroBuild open implementation · preview Specialized factory Idea → orchestrator → specialist artifacts → isolated process sandbox → test/fix loop → deploy-ready output STOPS AT Generated repository with PRD/design/source/test/deploy artifacts and local preview → Repository reviewer, release authority and independent production verifier GOOD Making the full-SDLC role story, artifact names, shared context patches, channel inputs and multi-provider adapters visible in a compact open implementation STOP Proving production readiness, semantic acceptance, safe deployment, recovery, security or the claim that no management is required HUMAN Plan confirmation, provider/credential scope, requirements, target deployment, semantic acceptance, release and recovery. no ⚒ ◐ ✓ △ ⌁ × ⚖ × E3 ↗ commercial operating Factory Software Factory commercial product · preview Issue → PR Private-preview lifecycle control plane STOPS AT Automation coverage/health state or workflow-specific run result → Configured downstream automation / software-team operator GOOD Manage and observe configured multi-stage Droid automations STOP Dashboard state is not execution or outcome truth; no public acceptance, rework or safe Day‑2 denominator HUMAN People configure stages, credentials, tools and permissions; repository and production policy still decide review, merge, release and remediation. planned ⚒ ● ✓ ◐ ⌁ △ ⚖ ● E2 ↗ commercial specialist IBM Bob Java Modernization commercial product · preview Legacy modernization Guided Java upgrade package STOPS AT Clean-compiling modified Java workspace with build summary → Java developer / modernization reviewer GOOD Real build-and-repair loops STOP Young evidence; build green only HUMAN Humans prepare a clean build/branch, approve plans and checkpoints, review tests and control PR/release. bounded ⚒ ◐ ✓ △ ⌁ × ⚖ ◐ E2 ↗ open workflow kit Mastra Factory open implementation · preview Issue → PR Typed issue-to-PR factory STOPS AT Reviewed PR with work item at terminal stage → Configured merge gate / repository authority GOOD Persistent typed handoffs STOP Local sandbox is unsafe HUMAN Persistent work-board transitions and configurable review/merge gates. planned ⚒ ◐ ✓ ◐ ⌁ × ⚖ △ E2 ↗ commercial operating Netlify Agent Runners commercial product · generally available Issue → PR Web preview runner STOPS AT Build run with Deploy Preview → Developer / publisher GOOD Netlify-native web changes and deploy previews STOP Web-delivery runner—not a general queue, verifier or Day‑2 operator HUMAN Internal Builder can propose; Developers retain deployment/publication authority. bounded ⚒ ● ✓ ◐ ⌁ △ ⚖ ● E2 ↗ open workflow kit Shopify Roast open implementation · preview Factory substrate Deterministic workflow engine STOPS AT Cog outputs, working-tree changes and author-defined grade → Workflow author or adopter CI / reviewer GOOD Replayable Ruby workflows STOP No secure execution plane HUMAN Humans author workflows, choose grades/checks, and control integration. no ⚒ ◐ ✓ ● ⌁ × ⚖ × E2 ↗ substrate prototype Thruwire Foreman open implementation · prototype Factory substrate Jev semantic supervisor over one coding worker STOPS AT Bounded run decision and worker-produced candidate with local run artifacts → Independent test/acceptance authority, then human repository and release owner GOOD An inspectable early design for turning model-estimated run state into bounded deterministic continue/steer/stop/retry/verify/escalate actions STOP Causal RCA, staging reproduction, application Day 2, production-grade durability, security isolation, autonomous release, or correctness authority HUMAN People define task and thresholds, control backend trust, adjudicate ambiguous completion, independently validate evidence, review/merge, and own release and production recovery. no ⚒ ◐ ✓ △ ⌁ × ⚖ × E2 ↗ open workflow kit Uncle Bob SwarmForge open implementation · preview Factory substrate Role-separated TDD swarm STOPS AT QA-reviewed worktree branch with test evidence → Human operator / repository reviewer GOOD Quality-heavy local construction STOP No enterprise authorization HUMAN Human observes/intervenes in tmux and owns acceptance. no ⚒ ● ✓ ● ⌁ × ⚖ × E2 ↗ reference reference Vercel Eve open implementation · preview Issue → PR Four-stage PR factory STOPS AT Draft PR after analyst, implementer and reviewer stages → Human who marks ready / repository merge authority GOOD Forkable analyst → implementer → reviewer pattern STOP Template lacks a qualified scheduler, governance and recovery plane HUMAN Human makes the draft ready and merges. no ⚒ ◐ ✓ ● ⌁ × ⚖ △ E2 ↗ internal Operating factory · internal only · production evidenced Stripe Minions Issue-to-PR maintainer
BUY / BUILD IT FOR High-volume, well-scoped maintenance in a deeply instrumented monorepo with curated internal tools. NEVER INFER OR DELEGATE Portable adoption, broad language conclusions, release ownership, and application operations. THE HUMAN WHO REMAINS Humans open/take over work, review every PR, and retain merge and release authority. FACTORY STOP → NEXT OWNER Human-review-ready PR with CI evidence → Stripe code reviewer / owning engineer
How to read the dots ● leading ● strong ◐ bounded △ thin × absent ? unknown Scores describe public evidence for this exact system—not vendor potential, model intelligence, or an averaged recommendation. WHY THE ROWS DIVERGE The bench gives the verdict. Now open the machines. Only after comparing the complete field do we descend into vendor product families. This prevents a public SDK, hosted worker, workflow dashboard or reviewer from lending capabilities to the rest of its logo.
THE MOST IMPORTANT COMPARISON CORRECTION
One company. Two different machines. Vendor families often bundle an executor, orchestrator, enterprise control plane and open adapters. Each assessed row now names the exact surface. Read across to see what you get, what it is good at, and the hard edge where a different owner takes over.
Runtime ≠ factory Factory.ai ⚙ GA commercial platform
Droid runtime
YOU GET Resumable CLI / SDK / JSON‑RPC coding sessions
✓ GREAT AT Embedding an agent in your own workflow
× HARD STOP Caller must supply durable orchestration, acceptance and release
ACCESS / PRICE $20–$200 individual plans; enterprise quote Primary source ↗ ▦ Private-preview lifecycle control plane
Software Factory
YOU GET Coverage/health dashboard + configured automations that dispatch Droid
✓ GREAT AT Managing lifecycle automations around Droid
× HARD STOP Dashboard is not the executor; downstream owners still accept artifacts and effects
ACCESS / PRICE Private preview; no public price Primary source ↗ Engine ≠ fleet OpenRewrite / Moderne ⚙ Apache‑2.0 specialist engine
OpenRewrite OSS
YOU GET One-repository semantic recipe diff + data tables
✓ GREAT AT Deterministic JVM/config transformations
× HARD STOP No enterprise campaign plane, cutover or live operations
ACCESS / PRICE Free OSS; recipe licenses vary Primary source ↗ ▦ GA commercial specialist factory
Moderne Platform + CLI
YOU GET Serialized-LST search and governed multi-repo campaigns
✓ GREAT AT Fleet-scale modernization and security work
× HARD STOP No runtime-equivalence oracle, release or application operations
ACCESS / PRICE Contract / license required; no public list price Primary source ↗ Non-inheritance rule Public SDK ≠ open runtime parser source ≠ supported recipe surface GA executor ≠ GA factory commercial fleet ≠ application operations PRODUCT-FAMILY ANATOMY · CAPABILITIES DO NOT FLOW SIDEWAYS
A logo is not a system. Open the machine. Select a family, then follow each independently assessed component. Interface, executor, computer, orchestrator, control room, verifier and production authority can come from the same vendor—or seven different owners. The exact artifact crossing each seam is the point.
Factory.ai 8 assessed components OpenRewrite / Moderne 8 assessed components GitHub agent ecosystem 7 assessed components GitLab Duo Agent Platform 7 assessed components Vercel AI SDK Factory / Eve 9 assessed components Warp Factories / Oz 8 assessed components Netlify Agent Runners 8 assessed components prompt / event / PR → Public adapters feed a proprietary executor; preview control surfaces manage automations; repository and production authority remain outside.
→ reviewer, release system or service owner Swipe the component chain → select any organ for its artifact, owner and hard edge.
↯ interface / adapter SDKs · Action · plugins public source · mixed licenses →
⚙ execution runtime Droid CLI · Exec GA · proprietary →
▣ compute / isolation Droid Computers GA managed or BYOM →
⌘ workflow orchestrator Missions · automations evolving / preview boundary →
▦ control / observability Software Factory private preview →
✓ verifier / evidence collector Droid Control · QA · review mixed GA surfaces →
⌁ live-operations investigator Incident Response private preview →
♙ external effect authority SCM · release · service owners customer-owned
↯ interface / adapter · public source · mixed licenses SDKs · Action · plugins Factory.ai
Primary source ↗ TERMINAL ARTIFACT / STATE typed call, event stream or CI/plugin request NEXT ACCOUNTABLE OWNER Droid runtime or calling application × CAPABILITY CLIFF Public wrappers do not expose or open the Droid executor. LICENSE / EDITION BOUNDARY Python SDK Apache‑2.0; public source elsewhere must be licensed component by component.
Family non-inheritance adapter ≠ executor executor ≠ orchestrator computer ≠ durable workflow dashboard ≠ verifier verifier ≠ release authority release ≠ recovered application THE FIRST DIFFERENCE · WHAT CAN YOU ACTUALLY ADOPT?
Five things called a factory. Five different bargains. Read one row across. Internal systems have the strongest company-specific proof but cannot be bought. OSS exposes mechanics but transfers operations to you. Managed products sell the control plane but hide it. References teach. Substrates supply parts.
Same question → ◈ Enterprise internal Observed—not adoptable Primary ↗ ▦ Composable substrate Scope—not ownership Primary ↗ ✓ Best at Deep company context, curated tools and the strongest real throughput evidence.
Inspectable mechanics, model choice and composable building blocks.
Hosted fleet, integrations, administration and vendor support.
Forkable architecture and a concrete starting grammar.
Reusable worker, workflow, sandbox or coordination machinery; a substrate may itself be OSS or managed.
× Terrible at Portable adoption, independent reproduction and transfer of its private context.
Turnkey enterprise safety, support, recovery or proof merely because source is visible.
Reproducible scheduler semantics, easy exit or independent replay of vendor claims.
Production maturity, security and operational outcomes from a template alone.
A complete outcome, acceptance oracle, release plane or technology department.
⚠ Hard edge Do not infer that an adjacent public SDK, agent or blog diagram is the internal factory.
Source availability is not a hostile-code sandbox, protected verifier or durable production control plane.
A contractual control is not source-visible behavior; edition, region and plan boundaries can change the actual system.
A reference implementation is evidence of a design—not a qualified multi-tenant service or support contract.
More agents and better scheduling can produce slop faster unless an independent authority can reject and recover effects.
↻ Application Day‑2 Mostly repository work. Ramp and Dropbox expose bounded telemetry-to-fix loops, still without production authority.
Usually keeps workers running—not the deployed application. Operations agents remain separate organs.
The vendor operates factory workers. Most products still stop before deployed-application operation and recovery.
No inherited application Day‑2. The adopter must build scheduling, policy, telemetry, release and recovery.
Worker/session recovery may be strong; application Day‑2 is generally outside the substrate.
♙ Human residue Company engineers own intent, monitors, review, merge, release and incidents.
Platform, security and service owners supply policy, credentials, qualification and recovery.
Customers retain intent, review, release, risk acceptance and vendor governance.
Adopters become the platform team and own every unstated production contract.
Architects choose the organs; operators own semantic completion, external effects and recovery.
You inherit You cannot inherit the factory; its scheduler, data, policy and integrations remain private.
Hosting, isolation, identity, secrets, upgrades, integrations, support and control-plane recovery.
Vendor availability, metered costs, data-residency choices, opaque upgrades and lock-in.
Hardening, durability, tenancy, authorization, support and the full operating model.
Composition, cross-component identity, handoffs, observability, effect closure and failure ownership.
✓ Best at Deep company context, curated tools and the strongest real throughput evidence.
× Terrible at Portable adoption, independent reproduction and transfer of its private context.
⚠ Hard edge Do not infer that an adjacent public SDK, agent or blog diagram is the internal factory.
↻ Application Day‑2 Mostly repository work. Ramp and Dropbox expose bounded telemetry-to-fix loops, still without production authority.
♙ Human residue Company engineers own intent, monitors, review, merge, release and incidents.
You inherit You cannot inherit the factory; its scheduler, data, policy and integrations remain private.
✓ Best at Inspectable mechanics, model choice and composable building blocks.
× Terrible at Turnkey enterprise safety, support, recovery or proof merely because source is visible.
⚠ Hard edge Source availability is not a hostile-code sandbox, protected verifier or durable production control plane.
↻ Application Day‑2 Usually keeps workers running—not the deployed application. Operations agents remain separate organs.
♙ Human residue Platform, security and service owners supply policy, credentials, qualification and recovery.
You inherit Hosting, isolation, identity, secrets, upgrades, integrations, support and control-plane recovery.
✓ Best at Hosted fleet, integrations, administration and vendor support.
× Terrible at Reproducible scheduler semantics, easy exit or independent replay of vendor claims.
⚠ Hard edge A contractual control is not source-visible behavior; edition, region and plan boundaries can change the actual system.
↻ Application Day‑2 The vendor operates factory workers. Most products still stop before deployed-application operation and recovery.
♙ Human residue Customers retain intent, review, release, risk acceptance and vendor governance.
You inherit Vendor availability, metered costs, data-residency choices, opaque upgrades and lock-in.
✓ Best at Forkable architecture and a concrete starting grammar.
× Terrible at Production maturity, security and operational outcomes from a template alone.
⚠ Hard edge A reference implementation is evidence of a design—not a qualified multi-tenant service or support contract.
↻ Application Day‑2 No inherited application Day‑2. The adopter must build scheduling, policy, telemetry, release and recovery.
♙ Human residue Adopters become the platform team and own every unstated production contract.
You inherit Hardening, durability, tenancy, authorization, support and the full operating model.
✓ Best at Reusable worker, workflow, sandbox or coordination machinery; a substrate may itself be OSS or managed.
× Terrible at A complete outcome, acceptance oracle, release plane or technology department.
⚠ Hard edge More agents and better scheduling can produce slop faster unless an independent authority can reject and recover effects.
↻ Application Day‑2 Worker/session recovery may be strong; application Day‑2 is generally outside the substrate.
♙ Human residue Architects choose the organs; operators own semantic completion, external effects and recovery.
You inherit Composition, cross-component identity, handoffs, observability, effect closure and failure ownership.
The non-transfer rule Private adoption ≠ portable product source visible ≠ production safe managed ≠ independently proven template ≠ service substrate ≠ factory FOUR GREEN LIGHTS THAT MEAN FOUR DIFFERENT THINGS
What exactly is allowed to say “go”? Scan left to right. Every row has a real green zone. The last column is the cliff: none may inherit application, data or user truth merely because its own object is green.
agent run → future merge → infra apply → release / runbook → independent outcome
● explicit ◐ bounded × not owned Marks describe the exact plane—not the vendor’s total feature list.
✓ GREEN EDGE Open durable execution and runtime governance for unattended Python agent workflows: leases, checkpoints, cost/tool/token/latency limits, approvals, model switching and auditable lifecycle actions. × FATAL EDGE No target-system transaction or independent application outcome. ♙ HUMAN RESIDUE Authors policy and target idempotency; verifies the external effect.
DAY‑2 Agent lifecycle only EDITION / PRICE Apache-2.0 backend + MIT SDK · Cloud early access · v0.5.0 preview Open full evidence fingerprint ↓ Composition rule: join these planes with artifact, target, principal, policy revision and source-native effect receipts. Never propagate a boolean “success” downstream.
THE AT-A-GLANCE MACHINE FINGERPRINT
Every factory has a different stopping point. Scan every enterprise, OSS, reference and substrate system in one wall. A role can be one more prompt, another model session, a process, a container or a VM. That says nothing about who owns work, what survives a crash, or who may say done . Unknown stays visible.
¶ Prompt Different role instructions inside one context. Separation is rhetorical. 0 ◌ Session Separate model contexts. Files, credentials and controller may still be shared. 1 ▥ Process Separate operating-system work, often on one host and trust boundary. 13 ▣ Container Filesystem/process boundary whose network, secrets and external effects still need policy. 8 ◇ VM Stronger worker isolation; it does not supply correctness, ownership or exactly-once effects. 7 ⌘ Mixed Multiple worker forms. Each path needs its own qualification result. 21 ? Black box The public evidence does not reveal the real execution geometry. 243
All systems · 293 Only source-visible · 53
● yes ◐ bounded × absent ? undisclosed
Factory / actual execution unit STATEDurable state CLAIMClaim / lease / fencing CODEWorkspace and execution isolation HANDValidated handoff contract WORKERWorker crash recovery CTRLController crash recovery DONEIndependent completion authority
internal Stripe Minions ◇ vm · pipeline around one coding worker? unknown ? unknown ● yes ◐ partial ◐ partial ? unknown ◐ partial internal Ramp Inspect ? black-box · Telemetry → reproduce → proposed fix? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown internal Uber Minion ? black-box · web / Slack / GitHub / CLI / API → background coding task → tested PR◐ partial ? unknown ? unknown ◐ partial ◐ partial ? unknown ◐ partial open Uber ADR ? black-box · Agent telemetry → normalization → triage → deep detector → security finding? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown internal DoorDash Flux ◇ vm · invocation surface → playbook → isolated cloud worker → reviewable result◐ partial ? unknown ● yes ◐ partial ? unknown ? unknown × no open DoorDash Agentic Orchestrator ▥ process · requirements → phase artifacts → isolated worktree implementation → antagonistic review → publish● yes ◐ partial ◐ partial ◐ partial ◐ partial ◐ partial ◐ partial internal Monzo Agent Chip ▣ container · event consumer → typed workflow / queue → isolated agent container → structured response◐ partial ? unknown ◐ partial ◐ partial ? unknown ? unknown × no internal Razorpay Slash ? black-box · Slack / ticket / CI → single-, batch- or multi-repo agent → tested PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown internal Duolingo AgentWorkflow Platform ⌘ mixed · registry / event → durable Temporal workflow → isolated coding worker → gated artifact● yes ◐ partial ◐ partial ● yes ◐ partial ◐ partial ◐ partial internal TRM AskNickiBot Factory ? black-box · Specification → cross-repository scaffold → human launch? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown internal Dexter Schwarm ▥ process · polling control plane → node process → tmux coding session● yes ◐ partial ◐ partial ◐ partial ◐ partial ◐ partial × no internal Harvey Spectre ? black-box · collaborative durable run → isolated worker → artifact / branch / PR● yes ? unknown ◐ partial ◐ partial ◐ partial ? unknown × no internal Microsoft Customer Zero Agent Platform ? black-box · Portfolio of lifecycle agents over GitHub and Azure? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown internal Spotify Honk / Fleetshift ▣ container · fleet scheduler → pod-scoped coding session → separate judge◐ partial ? unknown ◐ partial ◐ partial ? unknown ? unknown ● yes internal Dropbox Nova ? black-box · private shared execution layer around one coding session? unknown ? unknown ◐ partial ◐ partial ◐ partial ? unknown ● yes open OpenAI Symphony ▥ process · issue poller → isolated workspace → one coding worker◐ partial × no ◐ partial ◐ partial ◐ partial × no × no open GitHub Agentic Workflows ▣ container · event → read-only reasoner → buffered proposal → privileged broker◐ partial ◐ partial ◐ partial ● yes × no ? unknown ◐ partial internal Coinbase Mux ▥ process · human-driven parallel agent worktrees◐ partial ? unknown ◐ partial ? unknown ? unknown ? unknown ◐ partial open Shopify Roast ▥ process · declared sequential/parallel workflow× no × no × no × no × no × no ◐ partial internal Atlassian HULA / Rovo Dev ? black-box · Human-refined plan → code → evaluation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Block Goose + private G2 ▥ process · single persisted conversation with context-isolated subagents● yes × no × no ◐ partial ◐ partial × no ◐ partial internal Vercel AI SDK Factory ◇ vm · durable four-stage pipeline with bounded loops● yes ◐ partial ● yes ◐ partial ● yes ◐ partial ● yes commercial Warp Factories / Oz ⌘ mixed · hosted event-driven factory graph◐ partial ? unknown ◐ partial ● yes ◐ partial ? unknown ● yes reference Vercel Eve ◇ vm · durable sequential pipeline over Eve sessions● yes ● yes ◐ partial ● yes ● yes ◐ partial ● yes reference Vercel Open Agents ◇ vm · single durable agent with delegated subagents● yes ◐ partial ◐ partial × no ◐ partial ◐ partial × no substrate Deputies ⌘ mixed · durable queue and lease → worker → provider-specific sandbox session● yes ● yes ◐ partial ◐ partial ◐ partial ◐ partial × no substrate Thruwire Foreman ▥ process · bounded job → observe/coalesce events → Jev typed judgments → deterministic Python policy → continue/steer/stop/retry/verify/finish/escalate◐ partial × no ◐ partial ◐ partial ◐ partial × no ◐ partial open Fabro ⌘ mixed · explicit deterministic DAG● yes × no ◐ partial ◐ partial ● yes × no ● yes open Mastra Factory ⌘ mixed · persistent issue-to-PR workflow● yes ● yes ◐ partial ● yes ● yes ◐ partial ● yes commercial Tessl Agent ? black-box · Failures → skills/rules/evals PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Uncle Bob SwarmForge ▥ process · sequential role swarm with human-observable loops● yes ◐ partial ◐ partial ◐ partial ◐ partial × no ● yes open NTM / Agentic Coding Flywheel ⌘ mixed · operator-led blackboard/task-graph swarm● yes ● yes ◐ partial ◐ partial ● yes ◐ partial ◐ partial commercial Netlify Agent Runners ? black-box · Prompt → build → deploy preview? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Ona Software Factory ? black-box · 16 automations across build → deploy → monitor? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Factory Software Factory ? black-box · Lifecycle control plane + configured Droid automations? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Factory Droid runtime ? black-box · Interactive/headless agent session + SDK/JSON-RPC? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown substrate Claude Code Agent Teams ◌ session · interactive lead-and-peers team× no ◐ partial × no × no × no × no × no open Owain Lewis Factory ⌘ mixed · durable controller → leased worker attempts● yes ● yes ◐ partial ◐ partial ● yes ● yes × no substrate Agentic Factory ▥ process · append-first blackboard and baton protocol● yes ◐ partial × no ◐ partial × no × no ◐ partial substrate OpenHands ▣ container · event-stream agent runtime with delegated roles● yes ● yes ◐ partial ◐ partial ● yes ◐ partial ◐ partial reference StrongDM Factory / Attractor ? black-box · NLSpec graph pipeline + private scenario factory? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Elastic AI Software Factory ? black-box · 40+ compiled GitHub agent workflows? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial GitHub Copilot Cloud Agent ? black-box · Mission Control → ephemeral task → draft PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial GitLab Duo Agent Platform ? black-box · Agents + flows + runners + CI/security? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Cursor Cloud Agents ? black-box · Trigger → Firecracker VM → draft PR → CI follow-up? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Cognition Devin ? black-box · Task/API → fresh VM → change/review/PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Augment Cosmos ? black-box · Expert graph → workers → PR/review/dev verification? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Qodo Code Review ? black-box · PR + rules + history → ranked findings? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open PR-Agent ? black-box · Diff + prompt + artifacts → PR comments? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial CodeRabbit ? black-box · Incremental PR context → review → optional fix? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Greptile Agent + TREX ? black-box · Code graph + review swarm + sandbox tests? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Graphite Agent ? black-box · Stacked PR → review/chat/fix → merge queue? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Ellipsis Review Pipelines ? black-box · Parallel read-only reviewers → optional gatekeeper → comment? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Cursor Bugbot ? black-box · Parallel diff passes → dedupe/vote → findings/autofix? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown reference Vercel OpenReview ? black-box · One durable agent → review/edit/comment/push? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial AWS Transform ? black-box · Assess → transform → test → migration handoff? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial IBM watsonx Code Assistant for Z ? black-box · COBOL/PL/I service slice → Java + method validation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial IBM Bob Java Modernization ? black-box · Analyze → recipes → agent fix → build loop? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial GitHub Copilot Modernization ? black-box · Specialist agents: assess → plan → upgrade → Azure deploy? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Konveyor / MTA / Kai ? black-box · Static incidents → grounded fixes → re-analysis? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open OpenRewrite OSS ? black-box · One repository + semantic recipe → source diff/data tables? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Moderne Platform + CLI ? black-box · Serialized LST fleet + campaigns → patches/branches/PRs? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial CAST Imaging / Highlight ? black-box · Source → architecture graph → modernization findings? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial vFunction ? black-box · Static + runtime graph → service boundary → Java baseline? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Mechanical Orchard Imogen ? black-box · Capture behavior → generate slice → old/new equivalence → strangler cutover? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Google Mainframe Modernization + Dual Run ? black-box · Assess/rules → agent generation → Dual Run → cutover? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Stakpak Agent ? black-box · Alert/schedule → inspect infrastructure → execute bounded operation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open HolmesGPT ? black-box · Alert/question → telemetry investigation → RCA / optional remediation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open OpenSRE ? black-box · Incident evidence → competing hypotheses / test → scored RCA workflow? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Kubernaut ? black-box · Signal → AI analysis → governed workflow → effectiveness assessment? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Sympozium ? black-box · Policy CRD → ephemeral AgentRun Job/Pod → delegation / trace / memory? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Kagent ▣ container · Kubernetes-native per-agent runtime● yes × no ◐ partial ◐ partial ◐ partial ◐ partial × no open Akmatori ▣ container · incident controller → subagent investigation → proposal/monitor● yes ◐ partial × no ◐ partial ◐ partial ◐ partial × no open KubeRCA ⌘ mixed · alert intake → async evidence collection → RCA/history workspace● yes ◐ partial ◐ partial ◐ partial ◐ partial ◐ partial × no open Kubernetes AIOps Agent ▣ container · stable/canary evidence → analysis loop → promotion signal or GitHub handoff◐ partial ◐ partial ◐ partial ◐ partial ◐ partial × no × no commercial Sentry Seer ? black-box · Error/performance issue → telemetry + code RCA → patch / PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Datadog Bits AI ? black-box · Signal → hypothesis tree → investigation → Bits Code PR / bounded remediation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial New Relic Autopilot ? black-box · Alert → telemetry/knowledge investigation → reviewed remediation recommendation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Dynatrace Davis / Intelligence ? black-box · Grail/Smartscape problem → causal context → approved workflow action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Azure SRE Agent ? black-box · Azure evidence → response plan → governed mutation → verification? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial AWS DevOps Agent ? black-box · AWS alert → parallel investigation → RCA / mitigation proposal? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Gemini Cloud Assist Investigations ? black-box · Cloud evidence → parallel hypotheses → recommended fix? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown internal Google internal SRE AI ? black-box · Specialized agents → investigation / playbook → occasional mitigation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial PagerDuty SRE Agent + Automation Actions ? black-box · Incident agent → recommendation → separately authored runbook job? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial incident.io Investigations ? black-box · Parallel evidence / adversary → finding → human-reviewed PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Rootly AI SRE + MCP ? black-box · Incident evidence → draft finding/fix → human sign-off? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open NVIDIA NVSentinel ? black-box · GPU health event → policy action → node recovery / quarantine? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial NVIDIA Mission Control AHR ? black-box · Hardware trigger → diagnostic runbook → one-node repair → validation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open CloudNativePG ? black-box · Postgres desired state ⇄ observed cluster? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open midPoint ? black-box · Identity source → mapping → target reconciliation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Bytebase ? black-box · SQL proposal → checks/approval → locked rollout? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Dagster ? black-box · Asset/partition state → run/backfill → materialization evidence? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open StackStorm ? black-box · Event/rule → authored Orquesta workflow → action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Microsoft Sentinel playbooks ? black-box · Incident/alert/entity → Logic Apps playbook → connector action and run log? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Splunk SOAR ? black-box · Finding/case → approved asset action → playbook run and investigation record? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial TheHive 5 ? black-box · Alert/observable → case/tasks/analysis → responder/report/knowledge? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Cortex analyzers and responders ? black-box · Observable → analyzer/responder program → result or external action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Shuffle SOAR ? black-box · Webhook/schedule/manual trigger → visual workflow/app → output or action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open DFIR-IRIS ? black-box · Incident → collaborative case/evidence/task → report or module action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial BetterCloud IT Agent ? black-box · SaaS question → proposed action → human approval? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Torii Eko ? black-box · SaaS context → answer / insight? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Stitchflow ? black-box · IT policy → deterministic API/browser workflow → review → action report? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Corma Automation Builder ? black-box · Record admin task → generate plan → live test → approve → repeat? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Zluri ? black-box · HR/SSO event → connector-specific workflow → run log / manual fallback? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial 1Password SaaS Manager ? black-box · Integration/SSO/browser usage + contracts → reclaim/downgrade/deprovision workflow? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Lumos Albus ? black-box · Identity context → recommendation/policy → bounded deprovision path → review evidence? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Zylo Clarity AI ? black-box · Commercial SaaS graph → answer/recommendation → selected MCP record mutation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Nudge Security agents ? black-box · OAuth/browser-risk discovery → verdict/action → explicit rescan? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Josys AI Integration Builder ? black-box · Recorded admin browser activity → observational connector → platform workflow? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Iron Reins SaaS OSS assembly ? black-box · ERPNext finance + GLPI assets/contracts + midPoint identity + Baton adapters + bounded executor + missing PostgreSQL reconciler? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial ConductorOne ? black-box · Access request/review → connector-specific lifecycle action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Soda AI ? black-box · Metadata → draft quality contract/check → human approval? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Bigeye bigAI ? black-box · Monitor/lineage evidence → RCA and downstream impact? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Microsoft Graph TCM ? black-box · Snapshot/baseline → six-hour drift observation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Microsoft365DSC ? black-box · Tenant export → compare/test → PowerShell DSC convergence? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open CIPP ? black-box · Delegated tenant inventory/action → standards/JML/operator workflow? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial ServiceNow ITSM AI Agents ? black-box · Incident/service record → recommendation or narrow spoke action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial JSM Virtual Service Agent ? black-box · Question/intent → answer, triage or authored automation flow? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Microsoft Intune + Windows Autopatch ? black-box · Entra-enrolled device/app population → rings, policy, detection and update status? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown standard Apple Declarative Device Management ? black-box · Declaration/command → device sync → status and software-update phases? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Jamf Pro patch and managed updates ? black-box · Mac/mobile inventory → package/patch policy or managed OS update → status/log? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial BMC HelixGPT + ITSM ? black-box · Incident/change/service model → recommendation, collaboration or authored action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Atlassian Assets ? black-box · Objects/relationships → work-item context, impact and authored automation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Freshservice + Freddy AI ? black-box · Employee request → conversation/form → ticket, routing, update or resolution note? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open GLPI ? black-box · Inventory/assets/tickets/licenses → ITIL workflow, service catalog and impact view? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open NetBox ? black-box · Declared IPAM/DCIM/network model → automation input and topology context? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Snipe-IT ? black-box · Asset/license checkout → accountable possession and lifecycle record? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Argo CD ? black-box · Approved Git revision → rendered Kubernetes intent → continuous reconciliation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Flux ? black-box · Git/OCI/Helm source → controller graph → Kubernetes reconciliation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Kargo ? black-box · Discovered revisions → Freight → Stage → approved Promotion and verification? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Argo Rollouts ? black-box · Rollout spec → canary/blue-green exposure → AnalysisRun → promote or abort? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Flagger ? black-box · Canary policy → routed traffic → metric/webhook checks → promote or rollback? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Keptn ? black-box · Workload version → pre/post tasks and metric evaluations → lifecycle status? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown substrate OpenSLO ? black-box · Portable SLO document → implementation-specific validation and evaluation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Pyrra ? black-box · Typed SLO → recording rules, burn alerts, missing-metric alerts and UI? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Sloth ? black-box · SLO YAML/plugin chain → reviewable Prometheus rules and alerts? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Kayenta / Spinnaker ACA ? black-box · Control + experiment metric scopes → statistical classifications → weighted canary score? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Harness Continuous Verification ? black-box · Deployment + service instances + APM/log baseline → risk classification → pipeline action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Grafana SLO ? black-box · SLI ratio + target + long window → error budget, dashboard and burn alerts? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Nobl9 ? black-box · Cross-provider SLI → objective/budget history → alerts and reports? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Datadog Deployment Gates ? black-box · Service/environment/version + monitors or FDD → asynchronous pass/fail gate? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Breeze ? black-box · RMM/PSA + governed AI operator → typed action intent → endpoint execution? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Automox MCP ? black-box · Natural-language/MCP request → guarded Automox API call → asynchronous endpoint job? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Tanium Atlas / AEM ? black-box · Live Tanium endpoint graph → agentic plan/interface → RBAC-constrained package or deployment? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Pharaoh ? black-box · Fresh endpoint evidence → scoped expiring approval → Sentinel repair → technical check / escalation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Octofleet ? black-box · Endpoint inventory/control plane → polled Windows/Linux agent → command/package exit status? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Fleet ? black-box · Cross-platform inventory/policy/GitOps → deterministic endpoint API and queued work? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Tactical RMM ? black-box · Endpoint agent + policy scheduler → remote command, patch, inventory and alert execution? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open MeshCentral ? black-box · Self-hosted server + cross-platform agent → remote desktop, terminal and file session? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Rudder ? black-box · Central policy server + relays + agent → desired-state, hardening, patch and compliance convergence? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Uyuni ? black-box · Server + proxies + Salt minions → channels, packages, provisioning and action chains? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open osquery ? black-box · Cross-platform agent + SQL tables → scheduled queries and endpoint evidence? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Velero ? black-box · Kubernetes CRDs + plugins + object storage → resource/PV backup, migration and restore? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Kanister ? black-box · Blueprints + actionsets + CRDs → application-aware backup, restore and copy operations? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Stash ? black-box · Kubernetes operator + restic/CSI + addons → workload/PVC/database backup and restore? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Veeam Kasten ? black-box · Application discovery + policy/CRD/UI → backup, restore, DR, mobility and transforms? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial SAP LeanIX ? black-box · Portfolio evidence → TIME/6R insight → decision/transformation artifact? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Ardoq ? black-box · Graph + stakeholder evidence → rationalization candidate → roadmap? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial ServiceNow Enterprise Architecture ? black-box · CMDB/portfolio → scored disposition → approved demand or retirement project? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Bizzdesign Horizzon ? black-box · Architecture model → scenario/impact analysis → transformation design? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Essential Open Source ? black-box · EA ontology/import → generated portfolio and roadmap views? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown substrate Turbo EA ? black-box · Integrated EA repository → analysis/recommendation → initiative/task/connector? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Backstage Software Catalog ? black-box · Git/provider metadata → reconciled developer catalog → portal/actions? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Archi ? black-box · Human architecture model → ArchiMate diagrams/reports? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial ServiceNow SPM Demand Management ? black-box · Feedback/idea/demand → assessment/playbook → approved work → financial/benefit actuals? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Planview Portfolios + IdeaPlace ? black-box · Crowd ideas/formal demand → constrained scenarios → human-selected funded plan? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Aha! Ideas + Roadmaps ? black-box · Portal/integration feedback → reversible merge + scorecard → approval → roadmap? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Productboard + Pulse / Spark ? black-box · Source-linked feedback → insights/CIS/cited AI → product hierarchy and roadmap? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Jira Product Discovery ? black-box · Ideas + linked insights → formulas/views → Jira delivery link? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Enterpret ? black-box · 50+ feedback sources → identity/context graph + taxonomy → cited analysis/MCP? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open OpenProject ? black-box · Human-qualified work → work packages/roadmaps/costs → project delivery records? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Plane ? black-box · Intake → triageable work item → cycle/module/initiative → delivery dashboard? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Leantime ? black-box · Ideas/strategy/goals → milestones/projects → human-oriented execution? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Taiga ? black-box · Human backlog/epic/issue → Scrum or Kanban delivery? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Huly ? black-box · Docs/chat/CRM/tracker context → work/process records across a self-hosted collaboration suite? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open GitLab Community Edition ? black-box · Service Desk/email/API issue → repository-linked engineering work → CI/deploy flow metrics? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Fider ? black-box · Public/private feedback post + vote → canonical duplicate/status → public response? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown internal WorkOS Horizon / TARS ⌘ mixed · event / issue / observability signal → plan → isolated implementation or investigation → human-reviewed artifact◐ partial ? unknown ● yes ◐ partial ◐ partial ? unknown ◐ partial internal Google large-scale LLM migration system ? black-box · Nightly Kythe discovery → classify references → LLM edits → compile/test → developer and owner review? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown internal Meta SapFix + Getafix + Sapienz ? black-box · Crash/static finding → localize → generate and rank patches → isolated compile/crash/test validation → human approval? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown internal LinkedIn Quality Assurance Agent ? black-box · Approved natural-language journey → deterministic replay or VLM planning → action/effect evaluators → verified bug report? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Gas City ⌘ mixed · pack/formula → durable ready frontier → claimed agent session → gated work graph → merge slot● yes ◐ partial ◐ partial ◐ partial ● yes ◐ partial × no open SuperPlane ⌘ mixed · trigger → versioned canvas graph → component/runner → approval/channel → target integration● yes ◐ partial ◐ partial ● yes ◐ partial ◐ partial × no reference LangChain Open SWE ⌘ mixed · Slack / Linear / GitHub signal → persistent isolated task sandbox → agent/subagents → draft PR◐ partial ◐ partial ● yes ◐ partial ◐ partial ? unknown × no internal Zup CodeGen ▥ process · developer prompt → backend ReAct loop → WebSocket local tools → interactive result● yes × no × no ◐ partial ◐ partial ◐ partial × no commercial Darkshift (formerly FactoryKit) ◇ vm · ticket / spec / bug → isolated multi-repo worker → checks + skeptical reread + recorded browser QA → human-reviewed PRs◐ partial ? unknown ● yes ◐ partial ◐ partial ? unknown ◐ partial commercial Replit Agent ? black-box · Prompt/spec → cloud workspace → app artifact / deployable project? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open c9r Agent Orchestrator ▥ process · trigger / manifest → durable task graph → capability-matched shell agents → guards / approval → evidence and recovery● yes ◐ partial ◐ partial ● yes ● yes ◐ partial ◐ partial open Nexus Agents ▥ process · goal → route / execute through agent CLIs → adversarial vote and quality gates → audit → bounded routing feedback● yes ◐ partial ◐ partial ◐ partial ◐ partial ? unknown ◐ partial open BoundFlow ▥ process · governed workflow → leased operation → checkpoint / approval / next operation● yes ● yes ◐ partial ● yes ◐ partial ◐ partial × no open Zuul ⌘ mixed · review event → speculative cross-repo future state → voting jobs → report / merge● yes ● yes ◐ partial ◐ partial ● yes ● yes ◐ partial commercial Spacelift ⌘ mixed · VCS event / schedule → IaC plan → policy / approval → apply → optional drift reconciliation◐ partial ◐ partial ◐ partial ● yes ◐ partial ? unknown ◐ partial commercial Octopus Deploy ⌘ mixed · release / runbook snapshot → environment / tenant binding → queued steps and approvals → deployment record● yes ◐ partial ◐ partial ● yes × no ◐ partial ◐ partial open Rewind (anduaura) ? black-box · eBPF capture → incident snapshot → HTTP dependency replay → trigger comparison? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Kitaru ? black-box · Imported trace → worker-side replay → tool policy / evaluator → cohort result? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown research Causal Agent Replay ? black-box · Recorded trajectory → do(action/observation/context/policy) → sampled forward branch → user outcome function? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown research BranchLab ? black-box · Trace IR → replay/re-exec branch → policy/eval/causal diff → redacted evidence pack? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Kando ? black-box · Append-only event log → projected agent world → responders / budgets → fork, diff and strict replay? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Cake Agents ? black-box · Ticket/chat → private or shared Kubernetes session → OpenCode → tests/logs/artifacts → attributed commit or PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Runfusion Fusion ? black-box · Intent → planner → workflow graph → plan/review/execute/review → worktree/PR/merge? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open oncall-agent ? black-box · Momento incident → dedupe/queue → Teleport auth → Bedrock investigation → ranked hypothesis → optional GitHub PR → Slack report? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown reference Aegis ? black-box · Metrics feed → observe/diagnose/propose → Pomerium gate → admin mutation → live verification → JSONL event log? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown reference Software Factory v2 ? black-box · Spec → locked interface → AC-derived tests → implementation → mechanical gate → cross-family review → model jury → integration → outcome verification? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Config365 ? black-box · Tenant template → Git/Gitea desired state → WhatIf/approval → act_runner/PowerShell → Microsoft Graph? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open cat-factory ? black-box · Board/tracker → durable pipeline → per-run container → review/test/acceptance → PR/CI/merge? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open k8s-mechanic ? black-box · Cluster failure → stabilize/deduplicate → in-cluster OpenCode investigation → manifest validation → fix/investigation PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open FactoryOS ? black-box · Intent → spec → plan → tasks → bounded agent work → verification → approval? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Simple Long Horizon Agent ? black-box · Task → tools → plan/act/check loop → run trace → benchmark result? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Lvvphole Software Factory ? black-box · Signal → triage/planning → bounded executor → tests/review/security/docs → verifier → draft PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Super Simple Software Factory ? black-box · Target repository + skill → deterministic Python graph → bounded agent nodes → specs/code/tests/review/docs? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open EvoX Genesis ? black-box · High-level objective → recursive responsibilities → transient agents/worktrees → validation → accepted repository world? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown reference OurArk Genesis ? black-box · Trusted source body → inspect → descend/extend → inherited validation → independent lineage? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open OpenAgentsInc Agent IDE ⌘ mixed · intent → supervised work unit → replaceable worker / ACP session → evidence and review receipt● yes ◐ partial ◐ partial ● yes ◐ partial ? unknown ◐ partial open Keploy ? black-box · Observed traffic → dependency virtualization → replayed API/integration test? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open WireMock ? black-box · HTTP request → matcher/scenario → stub response or modeled fault? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial-plus-open-cli Okteto Environments ? black-box · Manifest/PR → isolated Kubernetes environment → tests/endpoints → cleanup? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial-plus-open-core vCluster ? black-box · Tenant cluster spec → virtual control plane → host/tenant sync → snapshot/sleep/restore? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Debezium ? black-box · Logical WAL → snapshot/watermark → schema-aware change event? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Kafka Connect ? black-box · Connector config → workers/tasks → durable offsets/status → rebalance? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown source-available PeerDB ? black-box · Peer + initial load → CDC flow → Temporal/state/retry → destination? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open pgcopydb ? black-box · Parallel snapshot → logical-decoding follow → catch-up → compare/cutover? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open pg_repack ? black-box · Table/index target → physical rewrite → short lock gates → rebuilt relation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open OpenClaw DataOps Guardian ? black-box · Alert → fresh evidence → allowlisted rollback → recovery reconciliation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Ravn Agents ? black-box · Detectors → signed template → typed actuator → post-check? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Pulse Agent ? black-box · Signal → route → diagnose → plan → gate → remediate → verify → postmortem? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown reference IRAS ? black-box · Alert → triage → context → RCA → plan → approval → apply → postmortem? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Reelier ? black-box · Approved skill → bounded effect → signed receipt → deterministic replay? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial OpenAI Codex Cloud ? black-box · Issue/task → isolated cloud workspace → background run → reviewable branch/PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Claude Managed Agents ? black-box · Session log → resumable harness → replaceable sandbox → outcome event / scheduled deployment? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Google Jules ? black-box · GitHub task → temporary Google Cloud VM → plan/code/test → branch/PR → optional schedule? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Salesforce Agentforce ? black-box · Agent Script/metadata → sandbox test → agent user → production org → CRM/external action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial SAP Joule / Joule Studio / AI Agent Hub ? black-box · Business intent → SAP context/Knowledge Graph → agent/app/workflow → governed runtime → ERP action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Workday Illuminate + Agent System of Record ? black-box · People/money context → role-based agent → ambient/delegate execution → Workday transaction → agent analytics? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Oracle Fusion AI Agent Studio ? black-box · Fusion business object/tool → agent team/app → role-gated action → monitored workflow result? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open ERPNext / Frappe ? black-box · Supplier/PO/invoice/payment/project record → role-scoped document workflow → queue/scheduler? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Odoo Community ? black-box · CRM/ERP app record → Python model/RPC/workflow → integrated business object? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open SuiteCRM 8 ? black-box · OAuth2/API record → ACL-aware Process API → hook/job/scheduler → CRM record or external request? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Zammad ? black-box · Ticket/user/org event → trigger/scheduler/AI action → webhook/API → audit and service record? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Dolibarr ? black-box · ERP/CRM record → REST/trigger/hook → business event or external integration? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Mem0 ? black-box · Conversation/tool episode → scoped fact extraction → SQL/vector/entity memory → filtered retrieval? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Graphiti ? black-box · Episode → entity/edge extraction → bi-temporal context graph → hybrid/MCP retrieval? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Zep ? black-box · Structured/chat episode → temporal graph + context lake → policy-bound agent context? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Letta / Letta Code ? black-box · Stateful agent session → editable core/shared blocks + archival memory/MemFS → later behavior? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Cognee ? black-box · Files/text/URLs/episodes → remember/ingest → graph enrichment → recall/improve/forget? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Steward ? black-box · Network discovery → graph inventory → mission/subagent → approval → device operation → verification? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Operant ? black-box · Slack/Teams human → OpenClaw gateway → per-user OAuth/policy → approved tool call → Postgres audit? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open OneCLI ? black-box · IdP → personal sandbox → Rust gateway credential injection → policy/approval → external tool? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open AgentPass ? black-box · Intent/evidence → policy/approval → idempotent action gate → provider receipt → verified observation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Siclaw ? black-box · Alert/question → isolated AgentBox → evidence/hypotheses/probes → RCA/recommendation? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open pgBackRest ? black-box · PostgreSQL/WAL → encrypted repository → verified restore/PITR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open WAL-G ? black-box · Database/WAL → compressed encrypted remote archive → fetch/PITR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Barman ? black-box · PostgreSQL fleet → central archive/WAL authority → recovery target? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open CloudNativePG + Barman Cloud Plugin ? black-box · Cluster desired state → PostgreSQL control plane → new-cluster recovery? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open OpenBao ? black-box · Identity/policy → secret/lease/PKI → audited access? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Atlas ? black-box · Desired schema → plan/lint → versioned migration → apply? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Flyway ? black-box · Versioned migrations/schema model → history/drift → fleet apply? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Liquibase ? black-box · Changelog/changeset → precondition/check → update/rollback history? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Skeema ? black-box · SQL schema in Git → lint/diff → PR → push? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open gh-ost ? black-box · MySQL/binlog → ghost table copy → lag/load throttle → atomic cut-over? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Cloudflare Agents ? black-box · Durable Object agent → schedule/queue/workflow → MCP or Code Mode → resumable state? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Cloudflare Sandbox ? black-box · Worker → isolated container → commands/files/processes/tunnels → artifacts? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Amazon Bedrock AgentCore ? black-box · agent session → microVM → governed tools/identity → trace and quota? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Foundry Hosted Agents ? black-box · container image → per-session sandbox → Entra identity → Responses/Invocations? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Google ADK ? black-box · agent/workflow graph → tasks/plugins/tools → sessions/artifacts/memory → eval? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Google Agent Runtime ? black-box · Python agent bundle → managed runtime → sessions/memory/eval → Cloud Trace? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Microsoft Intune Security Copilot agents ? black-box · Defender/Intune/Entra evidence → prioritized suggestion → admin-approved Intune action? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial NinjaOne Patch Intelligence / RMM ? black-box · Endpoint telemetry/vulnerability input → risk/policy → patch/script/reboot → activity and health status? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown commercial Atera RMM / IT Automation ? black-box · Agent telemetry → threshold/automation profile → patch/script/reboot → feedback report? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Microsoft Atlas ? black-box · YAML/JSON workflow → Azure DevOps/AD/ARM REST calls → logs/rendered outputs? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open ZeroBuild ? black-box · Idea → orchestrator → specialist artifacts → isolated process sandbox → test/fix loop → deploy-ready output? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown reference Agent-Native Runtime ? black-box · YAML agent graph → mesh-mediated tools/delegation → HITL/egress policy → replayable audit? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open ZeroClaw ? black-box · Channel/CLI → Rust agent loop → provider/tools/memory/runtime → schedule/subagent/audit? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Open Enterprise AgentOps Mesh ? black-box · Agent/use case → suitability/risk/data/eval/policy/security/audit/approval/readiness → launch candidate? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open AXON ? black-box · .ax declaration → parse/type/validate → Python MCP server or TypeScript → optional runtime/mesh governance? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Azure AgentOps Accelerator ? black-box · Agent config/dataset → eval/telemetry/Doctor → baseline comparison → evidence pack/CI workflow? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown reference Ona Software Factory / Memo ? black-box · Intent → private background automations → PR/review/incident loop → public Memo application? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown reference Microsoft SRE Agent public lab ? black-box · AKS break scenario → managed Azure SRE Agent → Azure Monitor/GitHub knowledge → investigation/remediation demo? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown reference Azure SRE Agent Demo Lab ? black-box · Bicep/PowerShell → AKS demo → seeded break → managed SRE Agent investigation → fix/readback? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open OpenGeni ⌘ mixed · API/webhook/UI → durable session → managed sandbox or connected machine → approval/artifact● yes ◐ partial ◐ partial ● yes ◐ partial ◐ partial ◐ partial open Polos ⌘ mixed · agent definition → durable workflow → sandbox/tool calls → approval/trace/artifact● yes ◐ partial ◐ partial ◐ partial ◐ partial ? unknown ◐ partial open Smithers ⌘ mixed · plain-English workflow authoring → agent-built JSX graph → watch/rewind/fork/replay → review/eval artifact● yes ◐ partial ◐ partial ● yes ◐ partial ? unknown ◐ partial open Agenhood ▣ container · fleet API/console → persistent agent container → task/workflow → snapshot/Git result◐ partial ◐ partial ◐ partial ◐ partial ◐ partial ? unknown × no open useAgent ◇ vm · web/Slack/REST/schedule → event-sourced thread → sandbox workstation → approval/artifact delivery● yes ◐ partial ● yes ● yes ◐ partial ◐ partial ◐ partial open Agent OS ⌘ mixed · project chat → plan/task graph → sandboxed coding agent → build/browser/vision verification → approval/deploy connector◐ partial × no ◐ partial ● yes ◐ partial × no ● yes open AgenticOps ? black-box · Alert/request → HealthIssue → RCA → risk-tiered plan → approval → AWS executor → resolve? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Kubernetes MCP Guard ? black-box · Kubernetes anomaly → bounded dry-run plan → OAuth/browser approval → digest-bound execution? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Xata Agent ? black-box · PostgreSQL signal → preset diagnostic/playbook → read-only finding → tuning proposal? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Cloud SRE Agent ? black-box · Log stream → threshold incident → triage/RCA/remediation → Go validator → PR/MR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Zelyo Operator ? black-box · CRD policy/scan → security/anomaly/drift correlation → validated YAML/IaC → GitOps PR? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open Cynative ? black-box · Security question → authorized read-only fanout → sandbox analysis → cross-checked finding? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open AOH ? black-box · Org/team/role/skill pack → validator/compiler → Hermes/Claude Code/Codex workspace? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open AgenticOS ? black-box · Versioned agent spec → governed runner → web/API/chat/automation → capability/MCP action → audit? unknown ? unknown ? unknown ? unknown ? unknown ? unknown ? unknown open OpenInspect / Open-Inspect ⌘ mixed · web / Slack / GitHub / Linear / schedule / alert / webhook → session sandbox → tests and artifacts → attributed draft PR◐ partial ◐ partial ◐ partial ◐ partial ◐ partial ? unknown ◐ partial ① WHAT ACTUALLY RUNS? Go workers launch Pi, Codex or Claude locally or on remote VMs, each in a managed git worktree.
Mixed: Multiple worker forms. Each path needs its own qualification result. ② WHAT SURVIVES? Runs, work items, attempts, heartbeats and publish state persist in SQLite.
Scheduler: Go control plane with SQLite state ③ WHAT CAN STILL LIE? It solves fleet mechanics unusually well but does not prove that the resulting code is correct, safe to release or operationally healthy.
Completion: Repository checks and humans remain the semantic oracle; the attached agent reports success. ● yes Durable state Runs, work items, attempts, heartbeats and publish state persist in SQLite.
● yes Claim / lease / fencing Leases, heartbeats, retries and reclaim are explicit source-visible mechanics.
◐ partial Workspace and execution isolation Worktrees and optional VMs isolate revisions; local host-user execution is not a hostile-code sandbox.
◐ partial Validated handoff contract Attempts, branch/ref/head-SHA and publish records are structured; semantic agent output is still self-reported.
● yes Worker crash recovery Heartbeats, lease expiry, retries and cancellation support recovery.
● yes Controller crash recovery Durable SQLite state lets the controller reconstruct work, subject to crash-window qualification.
× no Independent completion authority Repository checks and humans remain the semantic oracle; the attached agent reports success.
PUBLIC REPOSITORY ≠ PUBLIC FACTORY
What can you actually take home? The recent enterprise factories expose radically different public slices. Select one company: green is reusable or learnable, black remains inside the company, and red is the failure boundary that prevents capability inheritance.
Razorpay PUBLIC · NOT LICENSED AI Playbook + separate MIT payment MCP Duolingo APACHE-2.0 SHELL slack-ai-agent DoorDash APACHE-2.0 FACTORY Agentic Orchestrator v0.152.0 Uber OPEN ORGANS · PRIVATE FACTORY ADR + architecture and migration artifacts Monzo ARCHITECTURE ONLY No public implementation located ✓ PUBLIC SURFACE AI Playbook + separate MIT payment MCP Factory doctrine, skill contracts and readiness gates are inspectable. The Go MCP connector is reusable code—but defaults write-capable and lacks a server-side approval/compensation contract.
Primary: Playbook commit 2ccaa865 ↗ ≠ DO NOT INHERIT ▦ PRIVATE FACTORY Razorpay Slash Slash scheduler, worker isolation, Compass plugins, model policy, knowledge graph, merge gate and recovery controller.
Primary: Slash architecture ↗ × PRODUCTION CLIFF A Slash-authored DepGuard workflow self-merged, then failed at network, runner, hostname and branch-policy boundaries before a complete bot-authored rollback.
Inspect the hard edge ↗ The comparison rule: score the public component and the private factory separately. A shell does not inherit Temporal, a doctrine repository does not inherit a scheduler, an OSS sibling does not inherit Flux, and an architecture article does not become runnable code.
INTERNAL vs OSS vs MANAGED · SAME ROWS, NO MARKETING CATEGORIES
Compare fingerprints—not feature lists. The defaults expose three different ownership models. Change any selector; every column keeps the actual job, green zone, hard stop, Day‑2 boundary and human residue aligned.
Choose factories Default: internal · OSS · managed
internal · factory 1 Spotify Honk / Fleetshift Cognition Devin Google large-scale LLM migration system AWS Transform IBM watsonx Code Assistant for Z IBM Bob Java Modernization GitHub Copilot Modernization Konveyor / MTA / Kai OpenRewrite OSS Moderne Platform + CLI vFunction Mechanical Orchard Imogen Google Mainframe Modernization + Dual Run Stripe Minions Uber Minion DoorDash Flux Monzo Agent Chip Razorpay Slash Duolingo AgentWorkflow Platform TRM AskNickiBot Factory Dexter Schwarm Harvey Spectre Dropbox Nova OpenAI Symphony Vercel AI SDK Factory Vercel Eve Mastra Factory Netlify Agent Runners Ona Software Factory Factory Software Factory Factory Droid runtime GitHub Copilot Cloud Agent Cursor Cloud Agents WorkOS Horizon / TARS Meta SapFix + Getafix + Sapienz Darkshift (formerly FactoryKit) DoorDash Agentic Orchestrator GitHub Agentic Workflows Coinbase Mux Shopify Roast Warp Factories / Oz Thruwire Foreman Fabro Uncle Bob SwarmForge NTM / Agentic Coding Flywheel Owain Lewis Factory Elastic AI Software Factory GitLab Duo Agent Platform Augment Cosmos Sympozium Gas City SuperPlane LangChain Open SWE Ramp Inspect KubeRCA Kubernetes AIOps Agent Sentry Seer Datadog Bits AI New Relic Autopilot Dynatrace Davis / Intelligence Atlassian HULA / Rovo Dev StrongDM Factory / Attractor open · factory 2 Spotify Honk / Fleetshift Cognition Devin Google large-scale LLM migration system AWS Transform IBM watsonx Code Assistant for Z IBM Bob Java Modernization GitHub Copilot Modernization Konveyor / MTA / Kai OpenRewrite OSS Moderne Platform + CLI vFunction Mechanical Orchard Imogen Google Mainframe Modernization + Dual Run Stripe Minions Uber Minion DoorDash Flux Monzo Agent Chip Razorpay Slash Duolingo AgentWorkflow Platform TRM AskNickiBot Factory Dexter Schwarm Harvey Spectre Dropbox Nova OpenAI Symphony Vercel AI SDK Factory Vercel Eve Mastra Factory Netlify Agent Runners Ona Software Factory Factory Software Factory Factory Droid runtime GitHub Copilot Cloud Agent Cursor Cloud Agents WorkOS Horizon / TARS Meta SapFix + Getafix + Sapienz Darkshift (formerly FactoryKit) DoorDash Agentic Orchestrator GitHub Agentic Workflows Coinbase Mux Shopify Roast Warp Factories / Oz Thruwire Foreman Fabro Uncle Bob SwarmForge NTM / Agentic Coding Flywheel Owain Lewis Factory Elastic AI Software Factory GitLab Duo Agent Platform Augment Cosmos Sympozium Gas City SuperPlane LangChain Open SWE Ramp Inspect KubeRCA Kubernetes AIOps Agent Sentry Seer Datadog Bits AI New Relic Autopilot Dynatrace Davis / Intelligence Atlassian HULA / Rovo Dev StrongDM Factory / Attractor commercial · factory 3 Spotify Honk / Fleetshift Cognition Devin Google large-scale LLM migration system AWS Transform IBM watsonx Code Assistant for Z IBM Bob Java Modernization GitHub Copilot Modernization Konveyor / MTA / Kai OpenRewrite OSS Moderne Platform + CLI vFunction Mechanical Orchard Imogen Google Mainframe Modernization + Dual Run Stripe Minions Uber Minion DoorDash Flux Monzo Agent Chip Razorpay Slash Duolingo AgentWorkflow Platform TRM AskNickiBot Factory Dexter Schwarm Harvey Spectre Dropbox Nova OpenAI Symphony Vercel AI SDK Factory Vercel Eve Mastra Factory Netlify Agent Runners Ona Software Factory Factory Software Factory Factory Droid runtime GitHub Copilot Cloud Agent Cursor Cloud Agents WorkOS Horizon / TARS Meta SapFix + Getafix + Sapienz Darkshift (formerly FactoryKit) DoorDash Agentic Orchestrator GitHub Agentic Workflows Coinbase Mux Shopify Roast Warp Factories / Oz Thruwire Foreman Fabro Uncle Bob SwarmForge NTM / Agentic Coding Flywheel Owain Lewis Factory Elastic AI Software Factory GitLab Duo Agent Platform Augment Cosmos Sympozium Gas City SuperPlane LangChain Open SWE Ramp Inspect KubeRCA Kubernetes AIOps Agent Sentry Seer Datadog Bits AI New Relic Autopilot Dynatrace Davis / Intelligence Atlassian HULA / Rovo Dev StrongDM Factory / Attractor Public-evidence fingerprint 0–5; ? means unknown
Build 5 Proof 4 Day 2 1 Ent. 4 Open 0 Evidence 4
Build 4 Proof 5 Day 2 0 Ent. 1 Open 5 Evidence 2
Build 4 Proof 4 Day 2 2 Ent. 4 Open 3 Evidence 3
Primary job The main work pattern—not every feature it advertises.
Issue → PR One-shot issue → reviewed PR Factory substrate Role-separated TDD/architecture swarm Factory substrate Factory-as-code control plane Entity class What kind of system this exact assessed surface is; ownership is separate.
Operating factory Stripe · internal Workflow kit Uncle Bob / community · open Factory platform Warp · commercial Availability / maturity Whether this exact surface can be obtained and how far it has progressed are independent facts.
internal only production evidenced · Mature internal open implementation preview · Open local swarm commercial product preview · Early access Excellent zone The narrow reason to choose it.
High-volume, well-scoped maintenance in a deeply instrumented monorepo with curated internal tools.
A source-visible role-separated specification/TDD/QA option for teams that want explicit handoffs—especially Go, Clojure, and Java; no evidence it beats a single agent.
Multi-harness, multi-model cloud workers with event/schedule intake, human steering, and versioned skills.
Hard stop The inference or duty it cannot safely inherit.
Portable adoption, broad language conclusions, release ownership, and application operations.
Unattended cloud scheduling, enterprise authorization, broad language equality, and production operations.
Being called fully open or production-proven five days after launch; transparent audit of the managed scheduler/control room.
Application Day 2 Operating the deployed application, not keeping factory workers alive.
no Repository maintenance and CI repair, not deployed-service incident ownership. no Local software construction, not application operations. planned Monitoring-to-issue is in the design; broad production remediation is not demonstrated. Human residue The person who still supplies judgment, authority, or recovery.
Humans open/take over work, review every PR, and retain merge and release authority.
Human observes/intervenes in tmux and owns acceptance.
Versioned checkpoints, session steering/handoff, human spec review and merge in the reference flow.
Anti-slop mechanism How the system tries to prevent plausible but weak output.
Conditional repository rules, targeted lint, selective CI, curated MCP tools, and at most two repair rounds.
Gherkin, red/green/refactor, mutation testing, CRAP/DRY, architecture review, property tests, final QA.
Specs, repository checks, browser/computer-use evidence, structured review, trusted workflow revisions, and review-skill PRs.
Open / paid boundary The source-visible part and the part you must trust or buy.
Proprietary internal factory. stripe/ai is adjacent, not Minions source.
Open local orchestration and tool ecosystem.
Open workers, SDKs, environments, skills, and examples; proprietary hosted orchestration and enterprise gates.
Language reality Observed or documented coverage, not runtime-image inference.
Strongest evidence: Ruby/Sorbet backend. Broader support is undisclosed.
Canonical mutation/CRAP/DRY packs: Go, Clojure, Java. Other languages are editable but lack equivalent factory packs.
Stock images: Node/TS, Python, Go, Rust, Java, .NET, Ruby, browsers; runtime presence is not competence proof.
Model reality Published use or controlled comparison—not a generic list of APIs it can call.
Undisclosed; no public controlled model matrix.
Roles can use Codex, Claude, Copilot, and Grok; no controlled best backend.
Warp Agent, Claude Code, Codex, Gemini CLI, Copilot/OpenCode paths; BYO inference at enterprise tier.
Outcome evidence What has actually been published, including the denominator that is still missing.
Stripe reports 1,000+ wholly Minion-authored, human-reviewed merged PRs per week; no reproducible factory artifact.
Runnable source and explicit language-specific quality tooling; Sep. 2026 author statements favor one-agent use personally, but neither those statements nor the repository provide a controlled comparison, accepted-production denominator, or benchmark.
Public reference factory and worker/source graph are substantial; independent quality and long-horizon evidence remain open.
Factory / mechanism Excellent at Hard stop Day 2 Build? Proof? Day 2? Ent.? Open? Evidence?
internal Stripe Minions Stripe · Mature internal One-shot issue → reviewed PR EXCELLENT AT High-volume, well-scoped maintenance in a deeply instrumented monorepo with curated internal tools. HARD STOP Portable adoption, broad language conclusions, release ownership, and application operations. no 5 4 1 4 0 4 − EXCELLENT AT High-volume, well-scoped maintenance in a deeply instrumented monorepo with curated internal tools.
TERRIBLE AT / DO NOT INFER Portable adoption, broad language conclusions, release ownership, and application operations.
HARD EDGES ⚠ Private implementation ⚠ Ruby/Sorbet is the strongest disclosed target ⚠ No public defect or cost denominator LANGUAGES Strongest evidence: Ruby/Sorbet backend. Broader support is undisclosed.
MODELS Undisclosed; no public controlled model matrix.
DAY 2 REALITY Repository maintenance and CI repair, not deployed-service incident ownership.
WHERE HUMANS REMAIN Humans open/take over work, review every PR, and retain merge and release authority.
HOW IT FIGHTS SLOP Conditional repository rules, targeted lint, selective CI, curated MCP tools, and at most two repair rounds.
OSS / COMMERCIAL BOUNDARY Proprietary internal factory. stripe/ai is adjacent, not Minions source.
PUBLIC EVIDENCE Stripe reports 1,000+ wholly Minion-authored, human-reviewed merged PRs per week; no reproducible factory artifact.
internal Ramp Inspect Ramp · Mature internal Telemetry → reproduce → proposed fix EXCELLENT AT The strongest public example of connecting production telemetry to a bounded code-repair loop. HARD STOP Unsupervised instrumentation, noisy alerts, production authority, and whole-estate incident command. bounded 5 4 4 4 0 4 + internal Uber Minion Uber · Mature internal Scoped task → asynchronous implementation → tested human-reviewed PR EXCELLENT AT Background implementation and repeatable maintenance inside Uber's enormous, deeply integrated repository and developer-platform estate. HARD STOP Being mistaken for Uber's entire software-factory portfolio, inheriting another organ's sandbox or verifier, or owning deployment, incidents, rollback and business recovery. no 5 4 1 4 0 4 + internal DoorDash Flux DoorDash · Production internal platform Reusable playbook → isolated cloud agent → visible result EXCELLENT AT Running high-volume, repeatable engineering workflows through fast microVM sandboxes, scoped internal tools and one playbook callable from many developer surfaces. HARD STOP Turning task volume into a correctness claim, proving exactly-once external effects, or inheriting autonomous merge, deploy and application-recovery authority. bounded 5 3 2 4 0 4 + open DoorDash Agentic Orchestrator DoorDash OSS · Early OSS v0.152.0 Research → design → roadmap → implementation → antagonistic review → publish EXCELLENT AT Making long-running feature work explicit and resumable through deterministic phases, persistent artifacts, fresh contexts, bounded retries and independent-looking review roles. HARD STOP Being treated as Flux's public controller, a benchmark-proven autonomous engineer, a secure multi-tenant worker fleet, or a deployment and application-operations plane. no 3 3 0 2 5 3 + internal Monzo Agent Chip Monzo · Early production internal Slack / Linear / webhook → queued workflow → reviewed PR or incident evidence EXCELLENT AT Permissioned, productized background engineering tasks across many entry points, including bounded incident investigation and fix drafting inside Monzo's context. HARD STOP Portable adoption, broad autonomous deployment, independent incident truth, or treating likely-cause synthesis as verified RCA and remediation closure. bounded 5 3 3 4 0 4 + internal Razorpay Slash Razorpay · Production internal platform Slack / ticket / CI → single-, batch- or multi-repo agent → tested PR EXCELLENT AT Context-rich code generation, review and multi-repository work at substantial internal volume, with repository readiness and job-specific model evaluation treated as platform concerns. HARD STOP Assuming a humanless merge or Agent Ready score proves the change can run in its real environment, is behaviorally correct, or can recover the intended production outcome. bounded 5 4 1 4 0 4 + internal Duolingo AgentWorkflow Platform Duolingo · Production internal platform Registry / event → durable Temporal agent workflow → gated PR or structured result EXCELLENT AT Making narrow, reusable coding and support agents durable, observable, retryable and callable from Slack, internal sites, schedules and other workflows. HARD STOP Equating workflow durability with code correctness, compiling every target inside the worker, or granting agents unreviewed merge, deployment and incident authority. bounded 4 4 3 4 1 4 + internal TRM AskNickiBot Factory TRM Labs · Production internal platform Specification → cross-repository scaffold → human launch EXCELLENT AT Standardizing repeatable multi-repository engineering work behind shared retrieval, permissions, audit, evals and review conventions. HARD STOP Portable adoption, autonomous architecture or launch judgment, and causal attribution of broad organizational outcomes to the factory alone. bounded 4 3 2 4 0 3 + internal Dexter Schwarm Dexter · Early production internal Linear / schedule → reconciled worker → green PR EXCELLENT AT Small, well-scoped asynchronous tickets and scheduled maintenance where a live preview, CI and review feedback can close the repository loop. HARD STOP Large ambiguous redesigns, hostile multi-tenant execution, independent semantic proof, release ownership or deployed-application operations. no 4 4 1 2 0 3 + internal Harvey Spectre Harvey · Production internal platform Slack / web / schedule → durable isolated run → reviewable artifact EXCELLENT AT Collaborative, resumable cloud software work with explicit repository, tool, credential, history and artifact boundaries. HARD STOP Portable adoption, independent semantic verification, autonomous merge/release, or inferring production-remediation authority from Datadog access and incident investigation. bounded 4 3 2 4 0 3 + internal Spotify Honk / Fleetshift Spotify · Mature internal Fleet migration factory EXCELLENT AT Large, repetitive migrations where repositories have standard shapes, discoverable owners, and enforceable checks. HARD STOP Irregular codebases with missing context/tests; moving or operating the underlying data itself. no 5 5 1 4 0 4 + internal Dropbox Nova Dropbox · Production internal platform Validated monorepo session → branch EXCELLENT AT CI repair, flaky-test remediation, repeatable migrations and dependency breakage inside Dropbox's deeply curated monorepo. HARD STOP Portable adoption, autonomous architecture, release authority, or unsupervised production incident remediation. bounded 5 5 3 4 0 4 + open OpenAI Symphony OpenAI · Low-key preview / prototype Linear issue → workspace → Codex run EXCELLENT AT A small, inspectable scheduler specification for turning eligible issue-tracker work into isolated coding-agent runs. HARD STOP Being treated as a secure multi-tenant factory, durable exactly-once queue, independent acceptance system, release plane, or Day‑2 operator. no 3 2 0 1 5 3 + open GitHub Agentic Workflows GitHub · OSS public preview Event → read-only agent → safe-output broker EXCELLENT AT Auditable GitHub-native scheduled and event reasoning with a clear split between untrusted model work and privileged output application. HARD STOP Semantic truth, durable long-horizon product construction, autonomous merge/release, or deployed-application operations. no 3 4 0 4 5 4 + internal Coinbase Mux Coinbase · Mature internal Concurrency control for agent fleets EXCELLENT AT Letting engineers run many background coding agents across a large repository estate. HARD STOP Transparent verification, external reproducibility, Day‑2 operations, and causal claims from self-selected power users. no 5 ? 1 ? 0 3 + open Shopify Roast Shopify · Open workflow engine Deterministic + model workflow EXCELLENT AT Repeatable Ruby-first workflows and model-assisted migrations with replay and grading. HARD STOP Being mistaken for a secure cloud factory, deployment plane, or SRE system. no 3 4 0 1 5 2 + internal Atlassian HULA / Rovo Dev Atlassian · Deployed research precursor Human-refined plan → code → evaluation EXCELLENT AT Enterprise work where explicit human planning and code refinement are acceptable parts of the loop. HARD STOP Evidence transfer from the HULA study to current Rovo products or unattended operations. no 4 4 0 4 1 4 + internal Vercel AI SDK Factory Vercel · Production internal Classifier → analyst → implementer → reviewer EXCELLENT AT TypeScript-centric issue/backlog work with live product probes, independent review, and bounded learning loops. HARD STOP Application operations, broad polyglot claims, and treating its public Eve template as exact production source. no 5 5 1 3 2 4 + commercial Warp Factories / Oz Warp · Early access Factory-as-code control plane EXCELLENT AT Multi-harness, multi-model cloud workers with event/schedule intake, human steering, and versioned skills. HARD STOP Being called fully open or production-proven five days after launch; transparent audit of the managed scheduler/control room. planned 4 4 2 4 3 3 + reference Vercel Eve Vercel Labs · Open reference Four-stage issue → draft PR EXCELLENT AT Learning or forking Vercel's durable workflow + sandbox factory pattern. HARD STOP Turnkey enterprise adoption, application Day‑2, or production-factory equivalence. no 3 4 0 2 4 2 + substrate Thruwire Foreman Thruwire · 0.3.0 package metadata; untagged research prototype Jev semantic supervisor over one coding worker EXCELLENT AT An inspectable early design for turning model-estimated run state into bounded deterministic continue/steer/stop/retry/verify/escalate actions. HARD STOP Causal RCA, staging reproduction, application Day 2, production-grade durability, security isolation, autonomous release, or correctness authority. no 3 2 0 1 5 2 + open Fabro Fabro · Pre‑1.0 OSS Deterministic agent workflow graph EXCELLENT AT Explicit, auditable workflow-as-code with commands, conditions, parallelism, human nodes, checkpoints, and model routing. HARD STOP Turnkey enterprise identity/release/SRE or claiming non-Python competence from a SWE-bench run. no 4 5 1 2 5 4 + open Mastra Factory Mastra · Open alpha Typed persistent issue → PR EXCELLENT AT Source-rich TypeScript factory template with work boards, memory, traces, separate review, and human stage transitions. HARD STOP Untrusted code in LocalSandbox, complete release/monitoring ownership, or proven enterprise maturity. planned 3 3 1 2 4 2 + open Uncle Bob SwarmForge Uncle Bob / community · Open local swarm Role-separated TDD/architecture swarm EXCELLENT AT A source-visible role-separated specification/TDD/QA option for teams that want explicit handoffs—especially Go, Clojure, and Java; no evidence it beats a single agent. HARD STOP Unattended cloud scheduling, enterprise authorization, broad language equality, and production operations. no 4 5 0 1 5 2 + open NTM / Agentic Coding Flywheel Doodlestein ecosystem · Open composable ecosystem Multi-agent control + durable work graph EXCELLENT AT Operator-driven heterogeneous swarms, work ownership, recovery, code archaeology, and cross-platform release evidence. HARD STOP Zero-operator use, inheriting target-language quality from orchestration, or confusing release machinery with application SRE. bounded 4 4 2 2 5 3 + commercial Netlify Agent Runners Netlify · Hosted product Prompt → build → deploy preview EXCELLENT AT Web projects already on Netlify that benefit from isolated agents with build/deploy context and human publication. HARD STOP General enterprise software, private infrastructure operations, identity/data migration, or proof beyond Netlify's web boundary. bounded 4 3 2 4 1 2 + commercial Ona Software Factory Ona · Public 10-day experiment 16 automations across build → deploy → monitor EXCELLENT AT Seeing how cloud environments, scheduled/event agents, post-merge probes, and incident-to-issue loops compose. HARD STOP Proving enterprise replacement, mature Day‑2, low defect escape, or long-term maintainability from ten days and one app. bounded 5 4 3 3 2 3 + commercial Factory Software Factory Factory.ai · Commercial private preview Lifecycle control plane + configured Droid automations EXCELLENT AT Managing, connecting and observing triage, code generation, validation, release, documentation and incident-response automations. HARD STOP Serving as independently proven production automation, an inspectable control plane, or a generally available whole-SDLC replacement today. planned 4 3 2 4 0 2 + commercial Factory Droid runtime Factory.ai · Generally available commercial runtime Interactive/headless agent session + SDK/JSON-RPC EXCELLENT AT Embedding resumable coding-agent sessions in a CLI, application or external orchestrator, with MCP tools, policies, telemetry and optional managed computers. HARD STOP Owning a durable multi-stage factory, deciding semantic acceptance, or proving that a completed session produced the right business or production outcome. bounded 4 2 1 4 2 3 + open Owain Lewis Factory Owain Lewis · Developer preview Durable local/VM agent-fleet control plane EXCELLENT AT Scheduling many local or remote coding workers with durable runs, leases, retries, cancellation, worktrees, and a managed repository catalog. HARD STOP Proving the agents' code is correct, securely executing hostile repositories, releasing software, or operating deployed applications. no 3 2 0 1 5 3 + reference StrongDM Factory / Attractor StrongDM · Private factory + public specifications NLSpec graph pipeline + private scenario factory EXCELLENT AT A precise public design vocabulary for graph-structured pipelines, typed events, human gates, retries, checkpoints, model routing, and spec-following evaluation. HARD STOP Being reproduced from the public repositories: the private scenario corpus, digital twins, satisfaction evaluator, execution fleet, and promotion plane are absent. no 2 4 0 1 3 3 + open Elastic AI Software Factory Elastic · OSS workflow library 40+ compiled GitHub agent workflows EXCELLENT AT Narrow repository-maintenance agents with unusually explicit precision-over-recall, safe-output, evidence, no-op, and human-review design. HARD STOP Durable multi-run orchestration, application releases, production telemetry/remediation, or proving the reliability of the whole workflow library. no 3 4 0 2 5 3 + commercial GitHub Copilot Cloud Agent GitHub · Enterprise hosted task plane Mission Control → ephemeral task → draft PR EXCELLENT AT The strongest GitHub-native cockpit for assigning, steering, auditing, and reviewing bounded asynchronous coding tasks. HARD STOP Long-horizon or multi-repository work, self-hosted GitHub Server, releasing applications, or application Day‑2. no 4 4 0 5 0 3 + commercial GitLab Duo Agent Platform GitLab · GA enterprise DevSecOps substrate Agents + flows + runners + CI/security EXCELLENT AT Combining coding agents, custom flows, GitLab CI/security, runner execution, audit, composite identity, and self-hosted-model options in one DevSecOps surface. HARD STOP Secure-by-default background execution, open-source equivalence, autonomous deployment, or deployed-application incident/rollback ownership. no 4 4 0 5 2 4 + commercial Cursor Cloud Agents Cursor · Commercial durable PR factory Trigger → Firecracker VM → draft PR → CI follow-up EXCELLENT AT The best-disclosed durable commercial worker architecture: Temporal-backed runs, isolated microVMs, hibernation/resume, broad intake, and PR/CI follow-up. HARD STOP Multi-repository long-running work, mandatory independent acceptance, unattended deployment, or effect-verified application operations. no 5 3 1 5 0 4 + commercial Cognition Devin Cognition · Commercial coding/migration agent Task/API → fresh VM → change/review/PR EXCELLENT AT Large repetitive migrations and refactors that can be partitioned into bounded, independently testable work packages with enterprise administration. HARD STOP Indefinite factory processes, transparent durable replay, MFA/CAPTCHA-heavy workflows, or autonomous production remediation. no 5 3 0 5 0 3 + commercial Augment Cosmos Augment Code · Commercial programmable factory Expert graph → workers → PR/review/dev verification EXCELLENT AT The broadest assembled commercial factory here: versioned Experts, workers, PR repair/review, disposable dev deployment verification, and read-only incident investigation. HARD STOP Independent correctness when agents author and judge, transparent durability, permission consistency across surfaces, or autonomous live remediation. bounded 5 4 3 5 1 4 + commercial AWS Transform Amazon Web Services · GA commercial modernization suite Assess → transform → test → migration handoff EXCELLENT AT Broad, AWS-native .NET, mainframe, VMware, runtime and recurring code-estate modernization with documented unattended paths. HARD STOP Proving whole-system behavioral equivalence, owning target operations, or earning trust from a successful build alone. bounded 4 2 2 5 1 3 + commercial IBM watsonx Code Assistant for Z IBM · Commercial mainframe specialist COBOL/PL/I service slice → Java + method validation EXCELLENT AT Human-guided COBOL and PL/I transformation with unusually concrete captured-path and method-level equivalence tests. HARD STOP Autonomous mainframe replacement, exhaustive transaction/batch parity, data cutover, deployment, or target Day‑2 operations. no 4 4 0 5 1 3 + commercial IBM Bob Java Modernization IBM · Young commercial package Analyze → recipes → agent fix → build loop EXCELLENT AT Guided Java runtime, Java/Jakarta EE, WebSphere-to-Liberty, UI and dependency/CVE upgrades with a real compile/fix loop. HARD STOP Production-ready equivalence, fleet-scale maturity, autonomous deployment, cutover, rollback, or operations. bounded 3 2 1 3 1 2 + commercial GitHub Copilot Modernization GitHub / Microsoft · Commercial guided modernization Specialist agents: assess → plan → upgrade → Azure deploy EXCELLENT AT Inspectable multi-agent Java and .NET framework upgrades, including supported Azure rearchitecture and deployment paths. HARD STOP Unattended behavioral preservation, legacy-data parity, whole-fleet cutover, non-Azure neutrality, or runtime Day‑2. no 4 3 1 5 3 3 + open Konveyor / MTA / Kai Konveyor / Red Hat · Mature analysis + beta agent assistance Static incidents → grounded fixes → re-analysis EXCELLENT AT The strongest fully open Java modernization assessment and human-assisted remediation ecosystem in this cohort. HARD STOP Treating zero analyzer incidents as a successful build, preserved behavior, deployed system, or autonomous migration. no 2 2 0 3 5 4 + open OpenRewrite OSS OpenRewrite community / Moderne · Mature OSS transformation engine One repository + semantic recipe → source diff/data tables EXCELLENT AT Deterministic, semantically aware refactors and dependency/framework migrations in supported languages inside a repository build. HARD STOP Coordinating a governed enterprise fleet, recovering hidden business intent, proving runtime equivalence, migrating data/traffic, deploying, or operating applications. bounded 5 3 1 2 5 5 + commercial Moderne Platform + CLI Moderne · Generally available commercial platform Serialized LST fleet + campaigns → patches/branches/PRs EXCELLENT AT Searching, measuring and running governed deterministic modernization or security campaigns across hundreds or thousands of repositories. HARD STOP Inferring undocumented business semantics, proving runtime or data equivalence, performing release/cutover, or operating the changed applications. bounded 5 4 2 5 1 4 + commercial vFunction vFunction · Mature commercial decomposition platform Static + runtime graph → service boundary → Java baseline EXCELLENT AT Runtime-informed Java/.NET monolith understanding and Java service-extraction baselines with architectural observability. HARD STOP A complete automatic rewrite, .NET extraction parity, old/new behavior proof, autonomous cutover, or an APM replacement. bounded 3 3 2 4 0 3 + commercial Mechanical Orchard Imogen Mechanical Orchard · Closed platform with production case Capture behavior → generate slice → old/new equivalence → strangler cutover EXCELLENT AT The strongest publicly described behavior-first modernization loop, including production-data replay and side-by-side cutover evidence. HARD STOP Calling finite observed traffic exhaustive, generalizing beyond disclosed workloads, or inferring autonomous target operations. no 4 5 1 4 0 3 + commercial Google Mainframe Modernization + Dual Run Google Cloud · Commercial assembled pipeline Assess/rules → agent generation → Dual Run → cutover EXCELLENT AT An assembled mainframe modernization pipeline with a strong separate online/batch old-versus-new traffic comparison product. HARD STOP One autonomous factory, equal translation support for every assessed language, or post-cutover Day‑2 operations. no 4 5 1 5 1 3 + open Sympozium Sympozium AI · v0.9.5 · active development Policy CRD → ephemeral AgentRun Job/Pod → delegation / trace / memory EXCELLENT AT A Kubernetes-native agent operating plane for ephemeral runs, ensembles, delegation, MCP gating, sandbox hooks, NATS channels and optional memory. HARD STOP Treating AgentRun success, a trace or stored memory as causal diagnosis, target recovery or enterprise authorization. no 1 1 1 2 5 3 + open KubeRCA kube-rca contributors · Active OSS · 1.4.6 Alert → evidence collection → RCA workspace → feedback/history EXCELLENT AT A durable incident context room that joins Kubernetes, Prometheus, Tempo and optional logs with Slack/dashboard history, feedback and similar incidents. HARD STOP Being treated as a causal oracle, autonomous repair controller or proof that a retrieved historical remedy applies to the current fault. bounded 1 3 3 2 5 4 + open Kubernetes AIOps Agent Kevin Dubois / contributors · Active research prototype · no release tag Kubernetes evidence → multi-agent RCA → stable/canary decision → GitHub patch handoff EXCELLENT AT A concrete specimen for stable/canary evidence comparison, Kubernetes diagnostics and optional source patch/PR creation. HARD STOP Being trusted with rollout promotion or remediation because a model score, PR URL or nonempty response looks complete. bounded 1 2 2 2 5 3 + commercial Sentry Seer Sentry · Commercial AI debugging agent Error/performance issue → telemetry + code RCA → patch / PR EXCELLENT AT A direct, narrow error-to-code path with explicit root-cause, solution, code-change, PR-iteration and handoff stages. HARD STOP Being treated as a complete brownfield factory, independent reproduction environment or proof that a generated PR repaired production. bounded 2 3 3 4 0 4 + commercial Datadog Bits AI Datadog · Commercial suite with Preview surfaces Signal → hypothesis tree → investigation → Bits Code PR / bounded remediation EXCELLENT AT The broadest documented observability-to-code surface: investigation, source PRs, security/cloud/test findings, adaptive detection and bounded Kubernetes remediation. HARD STOP Being treated as one coherent autonomous SRE authority; its product-family breadth hides different permissions, previews, AI-credit and target-effect boundaries. bounded 2 3 4 4 0 4 + commercial New Relic Autopilot New Relic · Public Preview Alert → telemetry/knowledge investigation → reviewed remediation recommendation EXCELLENT AT A read-mostly operational assistant with explicit reasoning display, telemetry investigation, knowledge/memory context and human-reviewed rollback recommendations. HARD STOP Being treated as a source-code factory or autonomous IT operator; New Relic explicitly documents that source-code review is outside its intended design. bounded 0 3 3 4 0 4 + commercial Dynatrace Davis / Intelligence Dynatrace · Mature causal observability + Preview agentic actions Grail/Smartscape problem → causal context → approved workflow action EXCELLENT AT The strongest declared causal/topology substrate in this cohort, with live problem lifecycles, dependency context, affected-user estimates and approval-gated workflows. HARD STOP Being treated as an independent application-effect oracle; a correlated problem or successful workflow task does not prove full business recovery. bounded 1 4 4 5 0 4 + internal WorkOS Horizon / TARS WorkOS · Production internal; private Events, issues and observability signals → isolated background work → human-reviewed PR or finding EXCELLENT AT Context-rich event-driven work across Linear, Slack, GitHub, schedules and observability, with scoped identity and cloud sandbox separation. HARD STOP Portable adoption, inspectable recovery, or proving that a PR, self-improvement proposal or incident hypothesis safely changed production. bounded 5 3 2 4 0 4 + internal Google large-scale LLM migration system Google · Production internal; FSE 2025 paper Nightly Kythe discovery → classify references → LLM edits → compile/test → developer and owner review EXCELLENT AT Repeating a tightly specified cross-language migration across a very large monorepo while preserving a durable backlog of unprocessed locations. HARD STOP General modernization, unattended semantic design, irregular one-off migrations, production cutover or eliminating manual cleanup. no 5 4 0 4 0 5 + internal Meta SapFix + Getafix + Sapienz Meta · Historical production prior art (2018) Crash/static finding → localize → generate and rank patches → isolated compile/crash/test validation → human approval EXCELLENT AT Closing a narrow Android/Java defect loop from detected crash or null-dereference finding to an automatically validated patch candidate. HARD STOP Arbitrary feature work, broad RCA, multi-service production recovery, humanless deployment or evidence that the 2018 system remains available today. bounded 3 4 1 3 0 4 + open Gas City Gas Town Hall · v1.4.1 active OSS Customer Pack/formula → durable Beads work graph → replaceable agent sessions → gated merge EXCELLENT AT Open durable work ownership, configurable factory topology and heterogeneous agent/runtime orchestration without coupling work truth to one session. HARD STOP Turnkey semantic verification, hostile-code isolation, enterprise governance, shared managed runs or application operations. no 3 2 0 2 5 4 + open SuperPlane SuperPlane · v0.30.0 beta OSS + Cloud beta Trigger → versioned event/data-flow graph → runner/tool/approval → release action EXCELLENT AT Making integrations, agent/tool tasks, approvals and release workflows visible and composable in one source-visible control plane. HARD STOP Supplying semantic correctness, target-native transactions or compensation, resource-level least privilege, or generic application operations by itself. bounded 3 2 1 3 5 4 + reference LangChain Open SWE LangChain · Active MIT main; no releases Slack / Linear / GitHub → persistent task sandbox → agent/subagents → draft PR EXCELLENT AT A forkable open implementation of an organization's asynchronous coding agent, with real intake, identity mapping, sandbox adapters and draft-PR workflow. HARD STOP Protected validation, production access, exactly-once external effects, or inheriting Stripe/Ramp/Coinbase outcome evidence from architectural resemblance. no 4 1 0 2 5 4 + commercial Darkshift (formerly FactoryKit) Darkshift / Hashnode team · Managed product; public $99/user/month Spec, ticket or bug → isolated multi-repo cloud work → evidence-backed PRs EXCELLENT AT Turn bounded work into coordinated multi-repository PRs with cloud isolation, credential brokering, deterministic repository checks and recorded browser QA. HARD STOP Merge, release, production data change, application Day 2 or proving behavior beyond the selected checks and recorded browser path. no 5 3 0 4 0 4 + commercial Replit Agent Replit · Managed product Prompt/spec → cloud workspace → app artifact / deployable project EXCELLENT AT Fast browser-native application prototyping with an integrated workspace, preview and deployment path. HARD STOP Being treated as a governed enterprise software factory, causal incident investigator or independent production-recovery authority. no 4 2 0 3 0 3 + commercial Cake Agents Cake AI Technologies · Commercial cloud-coding workbench; public install example references chart 0.5.0; GA and production outcomes not independently established Ticket/chat → private or shared Kubernetes session → OpenCode → tests/logs/artifacts → attributed commit or PR EXCELLENT AT Regulated engineering teams that want customer-cloud coding sessions, repeatable tools, central provider/repository controls, and live human collaboration around feature tickets. HARD STOP A complete software factory, autonomous RCA, production release owner, or whole-IT Day-2 operator. Cake's separate website AIOps claims do not transfer to Cake Agents, and session tests/PRs do not prove application correctness or recovery. no 3 2 0 3 2 4 + open Runfusion Fusion Runfusion · MIT multi-node software-factory platform; active public development Intent → planner → workflow graph → plan/review/execute/review → worktree/PR/merge EXCELLENT AT Operating a visible multi-project coding-agent fleet with PostgreSQL state, worktree isolation, selectable workflows, model lanes, missions, mailbox, plugins and mobile/desktop/web control surfaces. HARD STOP Owning application production operations, causal RCA, data/identity/security recovery or safe autonomous merge merely because a task passed internal gates. bounded 4 3 3 3 5 4 + open oncall-agent Allen Helton · TypeScript/Bun AWS incident-response prototype Momento incident → dedupe/queue → Teleport auth → Bedrock investigation → ranked hypothesis → optional GitHub PR → Slack report EXCELLENT AT Making incident identity, short-lived AWS access, tool-scoped investigation, structured state transitions and PR-only remediation explicit. HARD STOP Being treated as a causal RCA oracle or direct production repair system; confidence and hypotheses remain model-produced and remediation is off by default. bounded 3 3 3 3 5 3 + reference Aegis Wally Ahmed · Loop Engineering hackathon governed on-call prototype Metrics feed → observe/diagnose/propose → Pomerium gate → admin mutation → live verification → JSONL event log EXCELLENT AT Testing denial feedback and re-planning: a tempting remediation is blocked, the reason becomes context, and a different action is attempted and checked against live telemetry. HARD STOP Generalizing a curated checkout service into autonomous SRE or assuming route-level authorization sees command-body risk. bounded 2 3 2 1 4 3 + reference Software Factory v2 hraedon · Archived Python/Regista contract-first factory reference Spec → locked interface → AC-derived tests → implementation → mechanical gate → cross-family review → model jury → integration → outcome verification EXCELLENT AT Contract-first handoffs, deterministic gates, failure routing back to contract revision and per-role/channel/model/prompt telemetry. HARD STOP Producing regulated production software, replacing brownfield systems or removing the final specification/outcome authority; those are explicit non-goals or human steps. no 3 4 1 2 5 4 + open Config365 Potsolutions · Self-hosted Microsoft tenant factory candidate Tenant template → Git/Gitea desired state → WhatIf/approval → act_runner/PowerShell → Microsoft Graph EXCELLENT AT A source-visible Microsoft configuration control plane that packages tenant records, Git-backed baselines, runner execution, approval, backups and explicit update paths. HARD STOP Proving complete Microsoft workload coverage, safe tenant convergence, user/data/security recovery or production maturity from a five-commit public GA channel. bounded 3 3 3 3 5 3 + open cat-factory Kibertoad · Self-hosted delivery control-plane candidate Board/tracker → durable pipeline → per-run container → review/test/acceptance → PR/CI/merge EXCELLENT AT A broad, source-visible factory substrate with runtime-neutral domain packages, Cloudflare/Node/local facades, runner images, SDKs, spend/RBAC, model/agent registries and a PR-centered delivery loop. HARD STOP Proving autonomous whole-IT, production application Day 2, business/data/identity recovery or causal RCA simply because it opens a green PR. bounded 4 4 3 3 5 4 + open k8s-mechanic Lenaxia · Kubernetes RCA-to-GitOps Day-2 specialist Cluster failure → stabilize/deduplicate → in-cluster OpenCode investigation → manifest validation → fix/investigation PR EXCELLENT AT A concrete, bounded failure-to-proposal loop with CRD state, severity, stabilization, read-only RBAC, secret/tool redaction, short-lived GitHub App tokens, dry-run and GitOps-only writes. HARD STOP General IT operations, causal certainty, business/user/data recovery, automatic production mutation or cross-platform model qualification. bounded 3 3 4 3 5 4 + open FactoryOS Milon Biswas · Repo-owned specification and workflow template Intent → spec → plan → tasks → bounded agent work → verification → approval EXCELLENT AT Keeping product truth, execution truth, workflow commands, agent skills and approval gates in version-controlled repository artifacts. HARD STOP Being mistaken for a hosted factory, worker scheduler, deployment system, identity plane, data plane or Day-2 operator. no 2 2 0 2 5 4 + open Simple Long Horizon Agent Simple Agent Lab · Readable Python long-horizon agent harness Task → tools → plan/act/check loop → run trace → benchmark result EXCELLENT AT A compact baseline for studying persistence, tool calls, verification, cost and model/harness differences without a large framework. HARD STOP Production-scale orchestration, enterprise identity, target effect, RCA, deployment, rollback or user support. no 2 3 0 1 5 4 + open Lvvphole Software Factory Lvvphole · Public v1.4 verifier-owned coding harness Signal → triage/planning → bounded executor → tests/review/security/docs → verifier → draft PR EXCELLENT AT A compact, unusually explicit example of putting lifecycle, governance and pass/fail authority in code around a replaceable coding agent. HARD STOP Durable unattended operations, production deployment, SaaS/identity/data actions, semantic RCA or treating static diff checks as independent correctness proof. no 2 3 1 2 4 4 + open Super Simple Software Factory Daniel Disler · Reusable stamped workflow kit Target repository + skill → deterministic Python graph → bounded agent nodes → specs/code/tests/review/docs EXCELLENT AT A minimal composition pattern that makes the deterministic controller own sequencing while coding agents remain bounded workers inside a repository. HARD STOP Hosted factory operation, durable fleet scheduling, enterprise identity, live deployment, rollback, data workflows or broad language support claims. no 2 2 0 2 5 4 + open EvoX Genesis EvoX / EMI Group · Long-horizon persistent software-world runtime High-level objective → recursive responsibilities → transient agents/worktrees → validation → accepted repository world EXCELLENT AT Persistent project state with disposable workers and recursive decomposition for extended software evolution rather than one long-lived model session. HARD STOP Enterprise application Day 2, identity/tenant/data authority, production deployment, target-side effects, business acceptance or assuming reported compiler/MESA demos generalize. no 3 3 1 2 4 3 + reference OurArk Genesis OurArk · Agent-body creation and lineage engine Trusted source body → inspect → descend/extend → inherited validation → independent lineage EXCELLENT AT Making an agent's behavior-defining repository, identity, private-state boundary and descent provenance explicit and human-custodied. HARD STOP Operating the descendant, autonomous IT, live remediation, production promotion or treating creation/recycle as a deployment system. no 2 3 1 2 5 4 + open Microsoft Atlas Microsoft · Public .NET workflow CLI / active-development repository YAML/JSON workflow → Azure DevOps/AD/ARM REST calls → logs/rendered outputs EXCELLENT AT Inspectable Azure-centric administration workflows with dry-run, retries, loops, secret redaction and no arbitrary shell execution. HARD STOP Being treated as Microsoft's autonomous IT department, a backup/restore system, an LLM factory, or an independent target-state/business-effect oracle. bounded 3 3 3 3 4 3 + open ZeroBuild PotLock · Public early-stage hierarchical software-factory repository Idea → orchestrator → specialist artifacts → isolated process sandbox → test/fix loop → deploy-ready output EXCELLENT AT Making the full-SDLC role story, artifact names, shared context patches, channel inputs and multi-provider adapters visible in a compact open implementation. HARD STOP Proving production readiness, semantic acceptance, safe deployment, recovery, security or the claim that no management is required. no 3 2 0 1 5 3 + reference Agent-Native Runtime agent-native-apps · Proof-of-concept policy-mesh and replay reference YAML agent graph → mesh-mediated tools/delegation → HITL/egress policy → replayable audit EXCELLENT AT Showing per-agent authority, data-context-aware egress review, bounded graph changes, human checkpoints and committed campaign logs as explicit runtime artifacts. HARD STOP Being a scalable production factory, distributed control plane, verifiable-credential system or independent business-effect oracle. bounded 2 5 2 2 5 4 + open OpenGeni CloudGeni · Early Apache-2.0 public organization runtime API/webhook → durable session → managed sandbox or connected machine → approval/artifact EXCELLENT AT A serious open control plane for durable, replayable agent sessions, organization workspaces, governed credentials, connected machines, memory and artifacts. HARD STOP Owning target-system truth, causal RCA, safe production changes, identity recovery or whole-IT Day 2; the repo explicitly leaves side-effectful execution and operator security review in scope. bounded 4 4 3 3 5 4 + open Polos Polos · Apache-2.0 public runtime; early adoption Agent definition → durable workflow → sandbox/tool call → approval/trace/artifact EXCELLENT AT A compact Python/TypeScript agent runtime that packages sandboxing, durable retries, approvals, cron/webhooks/events, Slack and OTel around user-authored agents. HARD STOP Being an independent correctness or RCA authority; cost-saving, replay and production claims are vendor claims until a matched target/effect trial. bounded 3 3 3 2 5 3 + open Smithers Smithers AI · MIT TypeScript workflow runtime; active public development Plain-English outcome → generated JSX workflow → agent/reviewer/eval → time-travelable run EXCELLENT AT A highly inspectable coding-workflow workbench with live watch, rewind, fork, replay, reviewer sandboxes, benchmarks, evals and multiple coding harnesses. HARD STOP Being a durable enterprise scheduler, target-effect oracle or application operator; time travel of workflow rows does not reverse Git, cloud or business side effects. bounded 4 4 2 2 5 4 + open Agenhood Appssemble · MIT self-hosted Docker agent fleet; early public project Fleet API/console → persistent agent container → task/workflow → snapshot/Git result EXCELLENT AT A straightforward self-hosted fleet with persistent workspaces, pluggable Claude/Codex/OpenCode drivers, schedules, workflows, SSE, Git backup, MCP and filtered egress. HARD STOP Being multi-region enterprise control, application Day 2 or safe autonomous remediation; its documented production path is a single VM/Docker Compose stack. bounded 3 3 3 2 5 4 + open useAgent useAgentHQ · AGPL-3.0 alpha coworker platform Web/Slack/REST/schedule → event-sourced thread → Linux workstation → approval/artifact delivery EXCELLENT AT An organization-facing coworker surface combining durable threads, isolated visible computers, typed gateway tools, Slack, knowledge/memory and native document artifacts. HARD STOP Being an enterprise-ready autonomous IT department; alpha APIs, commercial-license boundary, sandbox providers and source/effect semantics still need qualification. bounded 4 4 3 3 4 4 + open Agent OS earthwalker17 · Apache-2.0 local-first project OS prototype Project chat → task graph → sandboxed coding agent → build/browser/vision verification → gated delivery EXCELLENT AT The clearest new local end-to-end harness for turning model output into a verified project artifact through real builds, browser assertions, visual review and bounded recovery. HARD STOP Being a cloud factory fleet, enterprise IT operating plane or application Day 2 system; it is intentionally a one-builder local cockpit with external connectors. bounded 3 4 2 2 5 3 + open AgenticOps LiboMa · Public AWS operations project; self-reported closed-loop lab Alert/request → HealthIssue → RCA → risk-tiered plan → approval → AWS executor → resolve EXCELLENT AT Explicitly separating router, inventory, detector, RCA, plan, executor and reporting roles around AWS infrastructure. HARD STOP Independent RCA, cross-cloud production recovery, user/data/security closure and evidence of broad operational reliability. bounded 3 3 4 3 5 3 + open Kubernetes MCP Guard mirusser · Experimental .NET reference implementation Kubernetes anomaly → bounded dry-run plan → OAuth/browser approval → digest-bound execution EXCELLENT AT Making the approval, identity, plan digest, review digest and MCP mutation boundary concrete. HARD STOP General RCA, arbitrary Kubernetes operations, autonomous production changes or proving the workload recovered for users. bounded 3 4 3 3 5 4 + open Xata Agent Xata · Archived Apache-2.0 database specialist PostgreSQL signal → preset diagnostic/playbook → read-only finding → tuning proposal EXCELLENT AT Narrowing agent authority to source-native PostgreSQL diagnostics and explicitly avoiding destructive SQL. HARD STOP Broad IT operations, database mutation, migration, recovery, cross-system RCA or proving a user-facing fix. no 2 3 1 2 5 4 + open Cloud SRE Agent avivl · Public Go incident-to-PR pipeline Log stream → threshold incident → triage/RCA/remediation → Go validator → PR/MR EXCELLENT AT A clean ports-and-adapters separation with sanitized prompts, provider fallbacks, typed incident objects and a non-mutating PR delivery boundary. HARD STOP Live remediation, durable incident operations, arbitrary-language patch validation, target recovery or independent RCA. no 3 3 1 2 5 4 + open Zelyo Operator Zelyo AI · Early Apache-2.0 Kubernetes security operator CRD policy/scan → security/anomaly/drift correlation → validated YAML/IaC → GitOps PR EXCELLENT AT Combining Kubernetes/cloud security scanners, compliance, drift, cost and GitOps remediation in a source-visible operator. HARD STOP Causal RCA, direct production mutation, complete security closure, business recovery or proof that broad scanner coverage is complete. bounded 3 3 3 3 5 4 + open Cynative Cynative · Public read-only security-agent framework Security question → authorized read-only fanout → sandbox analysis → cross-checked finding EXCELLENT AT A deliberate read-only-by-construction cross-system security and drift investigation surface. HARD STOP Remediation, target changes, continuous fleet scheduling, complete source inventory or business recovery. no 2 4 1 3 5 3 + open AOH agenticdevops · MIT engine-neutral Ops harness MVP Org/team/role/skill pack → validator/compiler → Hermes/Claude Code/Codex workspace EXCELLENT AT Representing the IT department as portable roles, process skills, runtime requirements, model profiles and eval artifacts. HARD STOP Being mistaken for a scheduler, worker fleet, permission system, deployment authority or Day 2 operator. bounded 2 3 2 3 5 4 + open AgenticOS vstorm-co · Apache-2.0 self-hosted agent platform Versioned agent spec → governed runner → web/API/chat/automation → capability/MCP action → audit EXCELLENT AT Centralizing versioning, exposure, budgets, approvals, capabilities, knowledge, MCP and audit across many user-facing channels. HARD STOP Cross-system source truth, target-specific authorization, causal RCA, application recovery or whole-IT effect closure. bounded 3 3 3 4 5 4 + open OpenInspect / Open-Inspect Cole Murray · MIT self-hosted background-agent factory platform; source-qualified current main Signal or conversation → durable session → reproducible sandbox → coding harness → tests/artifacts → attributed draft PR EXCELLENT AT Providing a forkable session platform with multiple intake surfaces, per-session cloud development environments, provider choice, automations, multiplayer context, brokered secrets and reviewable repository output. HARD STOP Being treated as a complete organizational operating model, independent correctness oracle, multi-tenant SaaS, deployment authority or application Day-2 system merely because it can run agents and open PRs. bounded 4 3 2 3 5 4 + DEEP COMPARISON LABS Now isolate one mechanism at a time. The wall above answers which systems differ. The labs below explain why: write authority, deployment authority, telemetry judgment, Microsoft estate control, recovery, topology and verifier independence.
THE SLOP-TO-AUTHORITY MAP
Read-only agent. Write-capable system. Follow the artifact—not the marketing label. A model may hold no write token while a downstream broker publishes exactly what it asked for. Select any row to see where the factory earns trust and where its authority must stop.
untrusted need → reason → edit → prove → privileged broker → merge → deploy
Swipe the matrix → tap a factory → read its exact stopping point below.
Factory / control pattern Reason Edit Verify Publish Merge Deploy Fatal edge Primary
Dropbox Nova CI outside model ● yes ● yes ● yes ◐ bounded × no × no Publication identity, scheduler recovery and security are private; the human still decides whether code reaches production. ↗ OpenAI Symphony scheduler seed ● yes ● yes ◐ bounded ◐ bounded × no × no v0.0.2 has process-local claims and no protected independent acceptance or release plane. ↗ GitHub Agentic Workflows privileged broker ● yes ● yes ◐ bounded ● yes ◐ bounded × no The broker is still a writer. Compiler, sanitizer, policy or semantic-review failure can turn safe-looking output into a bad effect. ↗ Stripe Minions deterministic shell ● yes ● yes ● yes ● yes × no × no Private publication internals and human-reviewed PRs do not establish merge, release or live-service authority. ↗ Vercel AI SDK Factory separate reviewer ● yes ● yes ● yes ● yes × no × no The disclosed chain still stops before human readiness, merge, release and application operations. ↗ GitHub Copilot Cloud Agent branch-scoped worker ● yes ● yes ◐ bounded ● yes × no × no One repository, one branch, one PR and a 59-minute ceiling; the agent cannot approve or merge. ↗ ✓ GOOD AT Exact-commit sessions plus externally selected validation are excellent for repeatable monorepo toil.
× TERRIBLE AT / CLIFF Publication identity, scheduler recovery and security are private; the human still decides whether code reaches production.
♙ WHO TAKES OVER Nova hands a validated branch to review, CI/release policy and the accountable service team.
The rule: job green, PR opened, reviewer passed, merge complete and production healthy are five different claims. No factory inherits the next one.
THE PRODUCTION LAST-MILE MAP
Green deployment. Unknown outcome. Delivery is six separate jobs. Compare the controllers using identical columns, then select a row to see the exact claim it earns—and the attractive lie its green dashboard can still tell.
signed artifact → promote → converge → expose cohort → prove effect → recover
Swipe the matrix → tap a controller → inspect the exact stopping point.
Controller / actual job Artifact Promote Converge Expose Judge Recover Fatal edge Primary
Argo CD Git → Kubernetes reconciler ◐ bounded × no ● yes × no × no ◐ bounded Synced and Healthy are controller-specific state claims, not artifact provenance, user success, data correctness or safe rollback. ↗ Flux Headless Git/OCI reconciler ● yes ◐ bounded ● yes × no × no ◐ bounded The default reconciler authority is broad unless impersonation and tenant lockdown are configured; convergence still is not application truth. ↗ Kargo Artifact promotion orchestrator ◐ bounded ● yes ◐ bounded × no ◐ bounded ◐ bounded Manual approval bypasses verification/soak; generic metadata is outside Freight identity; controllers do not currently leader-elect. ↗ Argo Rollouts Progressive exposure controller ◐ bounded × no ◐ bounded ● yes ◐ bounded ◐ bounded No success/failure expression defaults a metric to success; stale, weak or misbound signals can promote a broken business effect. ↗ Flagger GitOps-native canary controller ◐ bounded × no ◐ bounded ● yes ◐ bounded ◐ bounded Metric thresholds inherit telemetry blind spots and cannot undo schema, queue, notification or other external effects. ↗ Helm / kubectl pipeline Direct imperative apply ◐ bounded ◐ bounded ◐ bounded × no × no ◐ bounded One-shot apply has no continuous ownership, fleet reconciliation, native progressive exposure or complete rollback transaction. ↗ ✓ EARNED GREEN ZONE Continuously compares approved Git-rendered intent with Kubernetes objects and reconciles a declared application boundary.
× FALSE-COMFORT CLIFF Synced and Healthy are controller-specific state claims, not artifact provenance, user success, data correctness or safe rollback.
♙ NEXT AUTHORITY CI/signing owns artifact identity; a rollout controller owns exposure; independent probes and service owners judge effects and recovery.
The failure that teaches the map: Git can be correct while the cluster is stale; the cluster can be healthy while the canary is misrouted; the canary can be green while users or data are wrong; a rollback can restore code while leaving irreversible effects behind.
THE JUDGE-OF-THE-JUDGE MAP
Green metric. Red reality. These systems are sold beside one another but do four different jobs. Read across identical columns: the yellow and red cells expose exactly where a plausible green result stops being evidence.
All 11 same hard-edge rows grammar defines the words accountant counts service health judge returns a release verdict actuator moves traffic / workload
exact artifact + actual cohort + complete fresh evidence + user · data · security · business + independent verdict + proved recovery
Swipe the matrix → tap a system → inspect its dangerous green default.
System / actual job Bind change Prove cohort Abstain Independent User/data/business Rollback Prove recovery Dangerous green default Primary
OpenSLO grammar · portable SLO schema × no × no × no × no ◐ bounded × no × no There is no runtime default: the specification does not collect, judge, gate or recover anything. ↗ Pyrra accountant · Prometheus SLO rule generator × no × no × no × no ◐ bounded × no × no Its absent-metric alert catches total absence—not partial loss, a wrong selector or a silently incomplete population. ↗ Sloth accountant · SLO → Prometheus compiler × no × no × no × no ◐ bounded × no × no Missing five-minute samples disappear from sum/count arithmetic; there is no distinct abstain state. ↗ Grafana SLO accountant · managed SLO accounting × no × no ◐ bounded ◐ bounded ◐ bounded × no × no New recording rules do not backfill; maintenance can create gaps, and the minimum objective window is seven days. ↗ Nobl9 accountant · commercial SLO platform × no × no × no ◐ bounded ◐ bounded × no × no No-data time-slice minutes count as good; no-data occurrence windows add neither good nor total. ↗ Keptn judge · Kubernetes lifecycle evaluator ◐ bounded × no × no ◐ bounded ◐ bounded × no × no No failure or warning criterion means an objective passes; nonblocking failures can leave the deployment successful. ↗ Kayenta judge · statistical canary judge ◐ bounded ◐ bounded × no ◐ bounded ◐ bounded ◐ bounded × no Caught metric errors are excluded from normal scoring; an all-error group can effectively score 100. ↗ Argo Rollouts actuator · progressive traffic actuator ◐ bounded ◐ bounded ● yes ◐ bounded ◐ bounded ● yes × no A metric with no success or failure expression defaults to success; a 2xx non-JSON response can also pass. ↗ Flagger actuator · GitOps canary actuator ◐ bounded ◐ bounded × no ◐ bounded ◐ bounded ● yes × no No values count toward failure, but there is no separate abstain state to distinguish missing evidence from harmful evidence. ↗ Harness CV judge · managed service verification ● yes ◐ bounded × no ◐ bounded ◐ bounded ● yes × no No Analysis passes unless Fail on no analysis is enabled; failure strategies can mark a verification failure successful. ↗ Datadog Gates judge · preview deployment gate ● yes ◐ bounded × no ◐ bounded ◐ bounded × no × no The official GitHub Action defaults fail-on-error to false; dry-run always passes and exhausted 5xx retries can succeed. ↗ ✓ WHAT IT ACTUALLY PROVES A vendor-neutral language for exchanging service-level objectives and alert conditions.
! DEFAULT THAT CAN LIE There is no runtime default: the specification does not collect, judge, gate or recover anything.
× WHAT IT CANNOT CLOSE A valid SLO document proves neither that observations exist nor that a release had the intended effect.
♙ REQUIRED HANDOFF A runtime evaluator, exact release/cohort ledger, protected policy and recovery authority must implement the contract.
The universal hard edge: none of these eleven independently binds the exact change to the actual population, proves complete fresh user/data/security/business outcomes, rejects ambiguity, causes safe rollback and proves whole-system recovery.
ONE MICROSOFT ESTATE · FIVE NON-EQUIVALENT VERBS
Seeing drift is not fixing it. Fixing it is not recovery. This is the operation-level lens the marketing categories hide. Select a row: green cells are source-backed authority, half cells are conditional, and red cells are a hard stop.
see → compare → change → keep changed → recover Each arrow needs a different contract.
System / actual role Observe Diff Apply Reconcile Restore Fatal hard edge Primary
Microsoft Graph TCM native observer ● yes ● yes × no × no × no Its only monitor mode is monitor-only; seven-day snapshots are not restore state. ↗ Microsoft365DSC open executor ◐ bounded ● yes ● yes ◐ bounded × no Completeness and auth vary per resource; apply is not tenant transaction or restore. ↗ CIPP open operations portal ◐ bounded ◐ bounded ● yes ◐ bounded ◐ bounded Core backup restores CIPP; scheduled backup restores only selected config—not content or a whole tenant. ↗ ServiceNow ITSM AI Agents managed coordinator ◐ bounded × no ◐ bounded × no × no The incident resolver writes recommendations; one spoke action is not estate authority. ↗ JSM Virtual Service Agent managed intake/flow ◐ bounded × no ◐ bounded × no × no Automation audit proves the flow ran—not that the downstream effect is correct. ↗ ✓ THE GREEN ZONE Cross-workload snapshots and property drift for documented resources.
× THE CLIFF Its only monitor mode is monitor-only; seven-day snapshots are not restore state.
♙ THE HANDOFF Administrator chooses another actuator and keeps external immutable evidence.
The universal rule: ticket complete, monitor green, export complete, workflow success and restore job success are all different claims. None inherits the next verb.
CONTINUOUS CHANGE / UPGRADE LAB
Seven organs. One chain nobody closes. These are organs—not full software factories. Most are excellent before merge and absent after deployment. Choose a lane to see the earned strength, the fatal edge, and the human authority that remains.
THE COMPLETE CHAIN detect → resolve → adapt → verify → stage / canary → reconcile / recover No single system closes all six. ● Strong ◐ Bounded × Missing Status means public evidence for this lane—not model potential.
CHANGE DOMAIN Detect Resolve Adapt Verify Stage / canary Reconcile / recover
⌁ Application dependencies Excellent PR factory; absent runtime operator Detect ● Strong registries + advisories Resolve ● Strong versions + lockfiles Adapt ◐ Bounded manifests, not behavior Verify ◐ Bounded inherits your CI Stage / canary × Missing stops at merge Reconcile / recover × Missing no deployed truth ⟳ Runtime / framework Can rewrite supported code; cannot own the cutover Detect ◐ Bounded metadata + recipes Resolve ◐ Bounded supported targets Adapt ● Strong semantic source rewrites Verify ◐ Bounded configured builds Stage / canary × Missing PR is the horizon Reconcile / recover × Missing no live rollback ▤ Schema / data DDL can succeed while the product fails Detect ● Strong schema diff + lint Resolve ● Strong ordered migrations Adapt ◐ Bounded DDL, some generated Verify ◐ Bounded ephemeral DB ≠ data Stage / canary ◐ Bounded expand / contract Reconcile / recover × Missing loss may be final ⇄ API / event contracts Sharp on declared shape; blind to undeclared meaning Detect ● Strong spec + schema diffs Resolve ◐ Bounded policy chooses break Adapt ◐ Bounded clients + schemas Verify ◐ Bounded structural contracts Stage / canary × Missing no traffic ownership Reconcile / recover × Missing no effect recovery ⬡ Kubernetes control plane / nodes / add-ons An upgrade DAG—not one green button Detect ● Strong deprecated APIs Resolve ◐ Bounded version matrices Adapt ◐ Bounded manifests + node plans Verify ◐ Bounded preflight ≠ workload Stage / canary ◐ Bounded rings per component Reconcile / recover × Missing control plane may not undo ◫ Image / GitOps Excellent convergence; health needs another organ Detect ● Strong registry policies Resolve ● Strong tag → digest Adapt ● Strong declarative commit Verify ◐ Bounded scan + CI only Stage / canary ◐ Bounded separate rollout Reconcile / recover ◐ Bounded Git truth ≠ health ☁ Vendor SaaS / model / API The provider can change reality without your commit Detect × Missing fragmented vendor feeds Resolve ◐ Bounded vendor chooses options Adapt ◐ Bounded client + config only Verify ◐ Bounded synthetic probes Stage / canary × Missing provider controls rollout Reconcile / recover × Missing often irreversible
EXCELLENT AT Keeping known package, image and infrastructure references current across a heterogeneous repository fleet, with explicit grouping and merge policy.
FATAL HARD EDGE A green or merged update PR proves neither production compatibility nor recovery. Renovate is policy-rich through conditional merge; Dependabot is GitHub-bound and PR-only; Updatecli propagation depends on verification you build.
HUMAN GATE Own required checks, review major and grouped updates, protect update-time credentials, approve promotion, and judge production rollback.
THE BALL-OF-MUD REALITY
Modernization is five organs—not one magic agent. Step through the Liferaft chain. Each station knows something different. The red card names the exact inference it cannot support.
⌕ Find the estate OpenGrok · Zoekt → ⌘ Resolve symbols SCIP · Kythe → ⟳ Transform deeply OpenRewrite · Coccinelle → ✂ Fill syntax gaps GritQL · ast-grep · Comby → ⚖ Prove the system protected tests · reconcile · shadow · canary · restore
THIS STATION CAN DO Map repositories, tokens, paths, history and observed cross-references at fleet speed.
THIS STATION CANNOT DO Text matches do not establish type identity, runtime behavior, business purpose or safe removal.
The missing organ Recovering undocumented business truth and proving preserved data, identity, timing, side effects and operations still require independent evidence plus accountable domain owners.
Primary: OpenGrok source ↗ THE REVIEWER TRAP
A second opinion is not automatically proof. Pick a reviewer. The picture shows how far its public evidence travels. A release-grade verifier needs a separate identity, model/context boundary, read-only review, clean runtime and protected oracle.
Qodo managed reviewer PR-Agent open reviewer CodeRabbit review + fix Greptile + TREX review + runtime evidence Graphite Agent review workflow Ellipsis review substrate Cursor Bugbot ensemble reviewer OpenReview self-review reference
✓ AUTHOR makes the change →
✓ REVIEWER different pass →
? SEPARATION identity + runtime →
? VERIFIER protected oracle
UNUSUALLY GOOD AT Multi-agent review, rules and broad SCM support
HARD EDGE Closed routing; vendor-run comment benchmark; no protected execution
Noise control balanced
Fix path suggests
Repository authority comment
Independence level 2/4 Open primary evidence for Qodo ↗ Iron Reins release rule AI review supplies evidence. Protected CI decides whether the exact artifact may progress. Canary effect decides whether production is healthy.
THE GREEN CHECK IS NOT THE FINISH LINE
Five stations. Five different claims. Choose a station. The yellow card says what it can prove; the red card shows the exact false-green it leaves behind.
◉ AI review Qodo · CodeRabbit · Greptile · Ellipsis → ⌕ Deterministic analysis OpenGrep · Semgrep · Sonar · CodeQL · Coverity → ⚗ Executable challenge protected tests · Hypothesis · mutation · fuzz + sanitizers → ◇ Artifact custody clean runner · reproducible build · SLSA · in-toto · Cosign · TUF → ✹ Production effect same-digest canary · business probes · telemetry · reconciliation · rollback
THIS STATION CAN SUPPORT A model found a plausible concern or improvement in the context it received.
DANGEROUS GREEN The reviewer shares the author's misconception, misses hidden behavior, or returns a neutral check.
Conjunctive release rule review evidence ∧ deterministic findings ∧ protected behavior ∧ artifact custody ∧ bounded production effect