95 software-factory candidates, plus 198 enabling systems. The System Atlas answers where a system sits. The Factory Map shows the public company patterns. This page is the decision surface: choose one family, compare three candidates, then pressure-test the Day‑2 boundary.
THE BRIDGE FROM ATLAS TO DECISION
First place the system. Then compare its finish line.
The Atlas answers where a system belongs in an IT department. This page answers what kind of work it repeats, what it can prove, and where humans still take over. Start with the job—not the logo.
The company map explained the patterns. This is where you choose one work family, put three candidates side by side, and inspect the exact handoff: artifact, proof, Day 2, human residue and evidence.
1 Hold the job constant2 Read the stopping point3 Ask who proves the result
01Choose a familysame kind of work→02Pick threeswap the columns→03Read the red edgewhere each stops→04Test Day 2does the loop close?
FIRST LENS · WHAT WORK REPEATSStart with product change factories.Ownership is a separate lens below, so a private Stripe loop is not accidentally compared as if it were the same thing as a hosted product.
ADVANCED LENSESFilter by ownership, mission or Day 2 evidence＋
SECOND LENS · WHO OPERATES THE CONTROL PLANE?
Compare the whole field only when you are intentionally mapping the ecosystem.
SHOW
23 in view
CHOOSE A COLUMN, THEN PICK A FACTORY
Selected column: #1
THE SIDE-BY-SIDE ANSWER
Same questions. Different stopping points.
Click a column to replace it, then scan the rows. A high build score never cancels a hard Day‑2 boundary.
Application Day 2Live service, not worker health.Day 2 · stops beforeRepository maintenance and CI repair, not deployed-service incident ownership.Day 2 · stops beforeMCP can expose operational tools, but the product has no governed release, runtime-monitoring, remediation, or rollback closure.Day 2 · stops beforeCloud Doctor/security experiments and CI follow-up produce PRs; no generalized deploy-observe-remediate-rollback transaction.
Human still needed
Humans open/take over work, review every PR, and retain merge and release authority.
The agent cannot mark its PR ready, approve, or merge; requesting users cannot approve their own Copilot-authored PR.
Normal terminal state is a draft PR; humans retain review, merge, deployment, and production authority.
Evidence
Stripe reports 1,000+ wholly Minion-authored, human-reviewed merged PRs per week; no reproducible factory artifact.
First-party constraints and audit behavior are strong; platform-wide cloud-agent task, merge, rework, and escaped-defect outcomes are not public.
Strong vendor architecture disclosure and large customer-reported job volumes; Cloud Agent contribution, defects, and causal outcomes are not independently isolated.
Evidence fingerprintpublic evidence, 0–5
Build5Proof4Day 21Ent.4
Build4Proof4Day 20Ent.5
Build5Proof3Day 21Ent.5
ORIENTATION APPENDIXSee the full family matrixUseful when you want the four clusters, public stopping point and Day‑2 counts on one screen.＋
THE MAP BEFORE THE TABLE
Four factory families. Four different finish lines.
Use this as the orientation layer. The cluster is the kind of work being repeated—not a vendor category. The stopping point is the furthest state public evidence supports, and the Day‑2 count is a warning against treating every “factory” as an operator.
family / repeated jobpublic edgehard questionDay 2 at a glancecandidates
ADVANCED QUALIFICATIONOpen RCA and protocol boundariesFor evaluating causal diagnosis, intervention authority and agent-to-tool handoffs.＋
RCA LENS · THE HARD PART OF DAY 2
A plausible cause is not a proven cause.
Place each system by its highest publicly evidenced RCA state. The ladder is cumulative: a system can have excellent telemetry and still stop before causal testing; a controller can execute a repair and still lack independent recovery proof.
Uses competing causes, required evidence and red herrings to test whether the diagnosis survives a falsifying question.
Hard edge: OpenSRE is strongest here as an evaluation plane; its public benchmark section reports no published results and it is not a production mutator.
Executes a declared workflow, records approvals/receipts and checks target-local health, alerts, metrics or desired state.
Hard edge: Those checks can miss users, data, security and external systems; Kubernaut’s alert-unavailable path must not be promoted as independent proof.
✓05
No public factory earns this across an enterprise estate
Independent recovery + recurrence
Separately proves source receipt, user journey, data invariant, security boundary, business outcome and recurrence prevention.
Hard edge: This is the Iron Reins acceptance target: evaluator outside the executor’s evidence path, explicit unknowns, compensation and a held recurrence window.
ADJACENT PROTOCOLS · READ BEFORE COMPOSING
Messages can travel. Authority does not.
These protocols make factories easier to connect across agents, editors, frontends and tools. They do not inherit the factory scheduler, tenant boundary, root-cause authority, source-native effect or Day‑2 proof. Treat each as a handoff surface and qualify it separately.
A2AA2A source ↗agent ↔ agentDiscovery, tasks, artifacts, streaming and push updates.A task status is not identity, durable scheduling, target effect or causal RCA.ACPACP source ↗editor ↔ coding agentSessions, files, terminals, permissions and resumable work.A completed session is not a reviewed commit, CI result, deployment or recovery.AG-UIAG-UI source ↗agent ↔ frontendEvents, state projection, tool progress and human-in-the-loop interrupts.A rendered state is not authorization, replay custody, source truth or effect proof.A2UIA2UI source ↗agent ↔ rendererDeclarative components, data binding and action surfaces.A component can render or emit an action without owning approval, identity or mutation.MCPMCP specification ↗agent ↔ tool / dataTyped tools, resources, prompts and connector context.A connector response is not source completeness, delegated authority or a native receipt.
WHAT TO DO NEXT
The comparison gives you a shortlist. Now test whether it can operate.
Use Day 2 to define the bar. Return to the System Atlas when you need the surrounding identity, data, platform and evidence organs; open Deep Research only when you need the underlying source wall.