# Enterprise and open software-factory field guide

Prepared: 2026-08-24  
Status: living evidence-normalized core cohort

## What this guide is for

This is the fast decision surface for Iron Reins. It does not ask which system has the most features or the most persuasive factory metaphor. It asks four questions:

1. What narrow job is this system unusually good at?
2. What hard boundary does it not cross?
3. Does it operate the deployed application after release, or only produce code and keep its own workers alive?
4. What public evidence justifies the answer?

This written core cohort contains 46 systems with a primary artifact and a falsifiable boundary. The interactive `/compare` wall is broader: Pass 058 projects all 54 explicitly admitted factory candidates plus eight adjacent review specialists through the same evidence grammar, with cohort filters and hard-edge search. It is not a claim that every private factory has been discovered. Private implementation, undisclosed language/model coverage and missing quality denominators remain unknown rather than receiving inferred credit.

## At-a-glance field guide

| Factory | Cohort | Excellent at | Hard stop | Application Day-2 | Public boundary |
|---|---|---|---|---|---|
| [Stripe Minions](https://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents) | private internal | High-volume, well-scoped maintenance in a curated monorepo | Not portable; no release or live-service ownership | No | 1,000+ reported merged PRs/week; implementation and defect/cost denominator private |
| [Ramp Inspect](https://engineering.ramp.com/post/why-we-built-our-background-agent) | private internal | Telemetry → reproduce → bounded repair proposal | Humans author monitors and retain merge/deploy/production authority | Bounded | Strongest public operational loop; duplicate and false-alert evidence is also public |
| [Spotify Honk / Fleetshift](https://www.engineering.atspotify.com/2025/11/spotifys-background-coding-agent-part-1) | private internal | Large repetitive fleet migrations with standard repository shapes | Failed on variable-context cohorts; does not operate migrated data | No | 1,500+ reported merged PRs; source private |
| [Coinbase Mux](https://www.coinbase.com/blog/coding-had-a-concurrency-problem-how-mux-helped-solve-it) | private internal | Concurrency control for many background coding agents | Verification internals, defects, review cost and Day-2 are undisclosed | No | 600+ users and 5,068 April-2026 PRs reported; power-user comparison is selection-sensitive |
| [Dropbox Nova](https://dropbox.tech/machine-learning/introducing-nova-our-internal-platform-for-coding-agents) | private internal | Validated CI repair, flaky-test work, repetitive migrations and dependency breakage in Dropbox's monorepo | Private scheduler/security/recovery; publication and production judgment remain outside the agent | Bounded | Roughly one in twelve PRs reported; no defect, cost, rework or risk denominator |
| [OpenAI Symphony](https://github.com/openai/symphony) | OSS preview/specification | Inspectable issue-to-isolated-Codex scheduling grammar | Process-local claims in v0.0.2; no stock durable lease, independent acceptance, release or Day-2 | No | Source-visible prototype and specification; live E2E tests opt-in and no outcome denominator |
| [GitHub Agentic Workflows](https://github.com/github/gh-aw) | OSS repository automation compiler | GitHub-native event reasoning with read-only agent jobs and brokered safe outputs | Broker/compiler/sanitizer still hold system write authority; not semantic proof or long-horizon factory durability | No | Source-visible v0.86.2 plus public hard-edge issues; no accepted-effect/defect denominator |
| [Shopify Roast](https://github.com/Shopify/roast) | OSS workflow engine | Ruby-first deterministic/model workflows, replay and grading | Not a secure cloud control plane, release system or SRE loop | No | Source-visible; longitudinal accepted-change evidence absent |
| [Atlassian HULA / Rovo Dev](https://arxiv.org/abs/2506.11009) | private/research | Human-refined enterprise planning, code and evaluation | HULA evidence cannot be transferred automatically to current Rovo products | No | Research study and outcomes public; implementation private |
| [Block Goose](https://github.com/aaif-goose/goose) + private G2 | OSS substrate + private factory | Provider freedom and extensible local tool use | Goose alone lacks factory isolation, acceptance, release and fleet orchestration | Unknown | Goose is source-visible; G2 is not |
| [Vercel AI SDK Factory](https://vercel.com/blog/building-a-software-factory-for-ai-sdk) | private internal + references | TypeScript issue/backlog work with independent review and live product probes | Four-week evidence window; not application operations or broad polyglot proof | No | Production system private; [Eve](https://github.com/vercel-labs/eve-software-factory-template) is a reference, not exact source |
| [Warp Factories / Oz](https://github.com/warpdotdev/warp-factory-examples) | open workers + hosted control plane | Multi-harness/model workers, factory-as-code and human steering | Managed scheduler/control room is not source-auditable; early-access evidence | Planned | Worker/SDK/example graph open; hosted orchestration commercial |
| [Vercel Eve](https://github.com/vercel-labs/eve-software-factory-template) | open reference | Learning/forking a durable four-stage issue → draft-PR pattern | Not turnkey enterprise, production-equivalent or Day-2 | No | Reference source visible; managed Workflow/Sandbox separately metered |
| [Vercel Open Agents](https://github.com/vercel-labs/open-agents) | open reference | Durable resumable Vercel-native cloud coding-agent substrate | No stock independent reviewer, release loop or broad language image | No | MIT reference; hosted primitives commercial |
| [WorkOS Horizon / TARS](https://workos.com/blog/project-horizon) | private internal | Event-driven context-rich work from issues, chat, schedules and observability in scoped cloud sandboxes | Private recovery; related Case source is not Horizon; no accepted-change/defect denominator | Bounded | 500 runs and 75 opened PRs reported; humans review/merge/release and operate |
| [Gas City](https://github.com/gastownhall/gascity/tree/v1.4.1) | OSS factory substrate | Durable work ownership separated from replaceable heterogeneous agent sessions | Trusted Packs, non-default fencing and adopter-built semantic gates; no application operations | No | MIT v1.4.1 source and candid failure fixes; no outcome denominator |
| [SuperPlane](https://github.com/superplanehq/superplane/tree/v0.30.0) | OSS + Cloud beta substrate | Versioned trigger, integration, runner, approval and release graphs | Beta; optional isolation and coarse RBAC; graph completion is not semantic/effect proof | Bounded | Apache-2.0 v0.30.0; Cloud beta limits and managed runner boundary separate |
| [LangChain Open SWE](https://github.com/langchain-ai/open-swe) | open reference factory | Forkable Slack/Linear/GitHub intake, persistent per-task sandboxes, subagents and draft PRs | Validation explicitly prompt-driven; no production access, releases or outcome denominator | No | MIT source on main; hosted sandbox/LangSmith dependencies separate |
| [Fabro](https://github.com/fabro-sh/fabro) | OSS pre-1.0 | Explicit deterministic workflow graphs, judges, humans and checkpoints | Enterprise identity/release/SRE are adopter work | No | Source and benchmark harness visible; complete-factory replication absent |
| [Mastra Factory](https://github.com/mastra-ai/softwarefactory-template) | open alpha | Typed persistent TypeScript issue → PR workflows with memory/traces | LocalSandbox shares host; release/monitoring vision exceeds shipped default | Planned | Alpha source visible; enterprise outcomes absent |
| [Tessl Agent](https://tessl.io/agent) | commercial improvement layer | Turning repeated failures into versioned skills, rules and evals | Not the executor, release authority or production operator | No | Portable artifacts/eval tooling visible; hosted control plane proprietary |
| [Uncle Bob SwarmForge](https://github.com/unclebob/swarm-forge) | OSS local swarm | Role-separated specification, TDD, mutation/complexity and QA workflows | Local/operator driven; language quality packs are uneven; no enterprise control plane | No | Strongest explicit packs are Go, Clojure and Java; no long-horizon outcome |
| [NTM / Agentic Coding Flywheel](https://agent-flywheel.com/flywheel) | OSS composable ecosystem | Heterogeneous swarm control, ownership, recovery, verification and release evidence | Operator skill and central acceptance remain essential; many components add burden | Bounded | Substantial component source; no single independent factory-level benchmark |
| [Netlify Agent Runners](https://docs.netlify.com/build/build-with-ai/agent-runners/overview/) | commercial hosted | Isolated web agents with Netlify build/deploy-preview context | Netlify/web boundary; not general enterprise software or private infrastructure operations | Bounded | Product docs public; control plane and full outcome denominator private |
| [Ona Software Factory](https://ona.com/stories/software-factory-what-we-learned) | commercial experiment | A visible build→review→deploy→monitor automation chain for one greenfield web app | Ten-day, self-authored experiment; taste/spec gaps and long-horizon stability unresolved | Bounded | Public experiment, app and report; platform/control plane commercial |
| [Factory Droid runtime](https://docs.factory.ai/droid-exec/overview) | GA commercial platform | Resumable interactive/headless agent execution through CLI, SDK and JSON-RPC | Caller still owns durable orchestration, semantic acceptance, release and production effects | Bounded | Public SDK/action/plugin adapters; core runtime and hosted control plane proprietary |
| [Factory Software Factory](https://docs.factory.ai/software-factory/overview) | commercial private preview | Managed lifecycle automation configuration, stage coverage and factory health/control-room state | Preview control plane; workflow executors and people still own the artifact and next effect | Planned | Product docs public; control plane proprietary; no public Software Factory price |
| [Claude Code Agent Teams](https://code.claude.com/docs/en/agent-teams) | experimental substrate | Interactive peer communication and parallel investigation | Not a durable unattended factory; coordination, resume and shutdown limitations are explicit | No | First-party docs; experimental and disabled by default |
| [Owain Lewis Factory](https://github.com/owainlewis/factory) | OSS fleet control plane | Durable local/remote worker scheduling, leases, retries and worktrees | Agent-reported semantic success; no hostile-code sandbox, release plane or application Day-2 | No | Source-visible developer preview pinned at `a9adbded`; no stable tag or outcome denominator |
| [Agentic Factory](https://github.com/jddelia/agentic-factory) | OSS coordination substrate | Durable batons, handoffs, messages, locks, review and verification records | Host runtime owns execution; recorded verification is not independent proof | No | Source-visible unreleased plugin pinned at `c65c174`; adoption unknown |
| [OpenHands](https://github.com/OpenHands/OpenHands/releases/tag/v1.15.0) | OSS coding-agent substrate | Persistent/forkable coding conversations, remote sandboxes, model breadth and benchmark adapters | Resume/cancellation do not make external effects exactly-once; no application Day-2 | No | Tagged MIT core/SDK/CLI and source-visible benchmark harness; enterprise outcome denominator absent |
| [StrongDM Factory / Attractor](https://github.com/strongdm/attractor) | private factory + open specifications | Graph pipeline, event, checkpoint, human-gate and model-routing specification vocabulary | Public repository is specifications, not the private runtime/scenarios/digital twins/evaluator | No | Apache‑2.0 specs, CXDB and AttractorBench public; claimed factory remains private |
| [Elastic AI Software Factory](https://github.com/elastic/ai-github-actions/releases/tag/v0.2.29) | OSS workflow library | Narrow repository agents with explicit silence/evidence/safe-output controls | GitHub Actions library, not durable factory or application-operations plane | No | Source visible; one 20-run triage audit recorded 0 success, and no estate-wide outcome denominator exists |
| [GitHub Copilot Cloud Agent](https://docs.github.com/en/copilot/concepts/agents/cloud-agent/about-cloud-agent) | commercial hosted | GitHub-native assignment, steering, audit, ephemeral execution and draft PRs | One repo/branch/PR and 59 minutes; no release or application Day‑2 | No | Strong product-boundary docs; global task/merge/defect outcomes unavailable |
| [GitLab Duo Agent Platform](https://docs.gitlab.com/development/duo_agent_platform/) | commercial/source-visible EE | Integrated agents, flows, runners, CI/security, audit and self-hosted model paths | Background sandbox/governance gaps; no autonomous deploy or application operations | No | Current source/docs expose architecture and durability fixes; effectiveness denominator missing |
| [Cursor Cloud Agents](https://cursor.com/blog/cloud-agent-lessons) | commercial hosted | Durable Temporal-backed microVM workers, broad triggers, draft PR and CI follow-up | Long-running excludes multi-repo; no independent acceptance or application operations | No | Strong architecture disclosure; customer outcomes vendor-hosted and not causally isolated |
| [Cognition Devin](https://docs.devin.ai/release-notes/2026) | commercial hosted/VPC | Large partitionable migrations with fresh VMs and enterprise controls | 30-day sessions, closed replay, human MFA/review/deploy, no production remediation | No | Migration case studies promising but incomplete/inconsistent; core closed |
| [Augment Cosmos](https://www.augmentcode.com/product/cosmos) | commercial factory | Expert graphs, PR repair/review, dev-namespace verification and incident investigation | Agent-correlated verifier, closed scheduler and human live-remediation authority | Bounded | Strongest commercial Day‑2 investigation evidence, but not effect-safe remediation |
| [Qodo Code Review](https://docs.qodo.ai/code-review) | commercial reviewer | Multi-agent review, organization rules and broad SCM/enterprise deployment | Closed routing and vendor-run comment benchmark; PR-Agent is a separate system | No | Strong transparent benchmark artifacts, but no protected execution or escaped-defect denominator |
| [PR-Agent](https://github.com/The-PR-Agent/pr-agent/tree/570f67ed5fc8db5be74c18df070bc20079b64b0d) | AGPL community reviewer | Auditable BYOM review experiments across many Git providers | Diff-local context, optional write surfaces and no mandatory executable oracle | No | Strong source evidence; public benchmark participation does not establish merge authority |
| [CodeRabbit](https://docs.coderabbit.ai/management/plans) | commercial reviewer/fix workflow | Incremental contextual review, learning, linked repositories and agent handoff | Closed engine; public skills are wrappers; comment acceptance is not correctness | No | Useful customer/adoption denominators; controlled defect/fix outcomes remain absent |
| [Greptile Agent + TREX](https://www.greptile.com/trex) | commercial semantic/runtime reviewer | Code graph plus generated sandbox tests and runtime artifacts | Generated tests may encode the same error; TREX is beta; no release oracle | No | Most important runtime-evidence movement in the specialist reviewer market |
| [Graphite Agent](https://graphite.com/docs/ai-reviews) | commercial GitHub review workflow | Stacked PRs, review/chat/fix and merge queue in one surface | Proprietary engine, GitHub-centric, no sealed execution and context-doc contradiction | No | Strong workflow evidence; benchmark snapshot showed precision/recall abstention extreme |
| [Ellipsis Review Pipelines](https://www.ellipsis.dev/docs/code-review) | commercial review substrate | Configurable models, read-only stages, isolated sandboxes, budgets and gatekeeper | Gatekeeper is optional; multiple model stages still are not protected truth | No | Best public independence machinery; product quality/outcome denominator still thin |
| [Cursor Bugbot](https://docs.cursor.com/bugbot) | commercial ensemble reviewer | Parallel passes, bucketing, voting, validators and cross-run deduplication | Closed models, neutral-by-default check and correlated cloud autofix | No | Large vendor-reported volume; resolution remains an LLM-derived proxy |
| [Vercel OpenReview](https://github.com/vercel-labs/openreview/tree/672deb21e70e471e0536d5ad7a67c14b8359e97e) | open reference workflow | Inspectable durable repository review and repair wiring | Same session reviews, edits and pushes with `--no-verify`; no protected schema/oracle | No | Valuable source-visible negative control for self-review claims |

## Score axes used by the interactive atlas

- **Build:** bounded intent → reviewable software change.
- **Proof:** independent checks, behavioral evidence and reviewer separation.
- **Day 2:** observes and repairs the deployed application, not merely its worker fleet.
- **Enterprise:** identity, policy, isolation, audit and administration evidenced publicly.
- **Open:** relevant factory machinery is source-visible; an open client does not make a hosted control plane open.
- **Evidence:** implementation, outcome denominator and longitudinal or independent support.

`?` means public evidence is insufficient. It does not mean zero capability.

## Non-transfer rules

- A reference template is not the source of a private production factory.
- A coding-agent substrate is not a software factory without intake, isolation, verification, integration and recovery machinery.
- Repository maintenance and CI repair are not application Day-2 operations.
- Worker restart/resume is factory self-operations, not ownership of the deployed service.
- Language runtime presence in an image is not evidence of language competence.
- Model availability is not evidence that one model is best for a repository/task class.
- Merged PR counts do not reveal defect escape, review burden, reverts, cost or causal productivity by themselves.
- OSS and commercial/enterprise editions are separate systems with separate evidence.

## Expansion gate

A candidate enters this table only after Iron Reins records:

1. a canonical first-party repository, paper, documentation page or engineering account;
2. the exact system boundary and whether the material control plane is public;
3. at least one narrow strength and one disconfirming hard edge;
4. Day-2 scope separated into factory uptime, repository maintenance and deployed-application operations;
5. known language/model evidence without inheriting capability from a shell or container;
6. the human authority boundary and slop-control mechanism; and
7. an evidence-strength judgment that preserves unknowns.