Repository maintenance and CI repair, not deployed-service incident ownership.
STEP 2 · FACTORY MAPSee what the big companies built.
See what the big companies built.
Then cluster the work.
Stripe, Spotify, Vercel, GitHub and the open builders are not building one universal “AI software factory.” This page owns the examples and the clusters: what goes in, what comes out, and where the public evidence stops. Comparison comes next.
THE SHORT ANSWER
A software factory is a repeatable production loop around a bounded change.
Some factories produce a PR. Some repeat migrations. Some operate a live domain. Some are only the worker substrate. The stopping point matters as much as the headline.
1What goes in?issue, signal, migration or request2What comes out?PR, patch, report, rollout or recovery3Who owns the next state?reviewer, deployer, operator or nobody yet
THE FACTORY MAP · AT A GLANCEThere is no single factory.
There is no single factory.
There are four loops.
THE BEGINNER ANSWER
A factory is a repeatable loop around one kind of change. Compare the loop it owns—not the size of the company or the ambition of the demo.
Read each card as: input → output → stopping point.01Name the inputissue, signal, migration or request→02Find the outputPR, patch, report, rollout or recovery→03Mark the edgeartifact, proof, live service or unknown→04Ask who staysreviewer, operator, vendor or nobody
⚒CLUSTER 01 · 23 ASSESSED
Product change factories
Issue-to-PR systems turn bounded product intent into code, tests and a reviewable change. Stripe, Vercel, GitHub and many managed agents live here.
PUBLIC EDGEreviewable changeIssue → code → PR / patch
0 evidenced12 bounded9 before
StripeStripe MinionsUberUber MinionDoorDashDoorDash FluxMonzoMonzo Agent Chip
Compare this loop →⟳CLUSTER 02 · 13 ASSESSED
Transformation factories
Migration and modernization systems repeat a known change family, where eligibility, rewrite rules and regression checks can be made explicit.
PUBLIC EDGEmigration artifactInventory → rewrite → regression report
0 evidenced5 bounded8 before
SpotifySpotify Honk / FleetshiftCognitionCognition DevinAmazon Web ServicesAWS TransformIBMIBM watsonx Code Assistant for Z
Compare this loop →✹CLUSTER 03 · 7 ASSESSED
Day‑2 operators
RCA, AIOps and rollout systems face the hardest seam: turning incomplete evidence into a safe, testable operational decision.
PUBLIC EDGEnarrow live loopSignal → hypothesis → bounded action
0 evidenced7 bounded0 before
RampRamp Inspectkube-rca contributorsKubeRCAKevin Dubois / contributorsKubernetes AIOps AgentSentrySentry Seer
Compare this loop →◇CLUSTER 04 · 17 ASSESSED
Factory substrates
Open runtimes, durable workflows and worker planes provide pieces you can assemble—but you inherit the proof, identity and operations layer.
PUBLIC EDGEworker / workflowRequest → session → handoff
0 evidenced3 bounded13 before
DoorDash OSSDoorDash Agentic OrchestratorGitHubGitHub Agentic WorkflowsCoinbaseCoinbase MuxShopifyShopify Roast
Compare this loop →DEEPEN THE MAPOpen the named company examples, stopping-point chart and rosterThe first screen is the teaching layer; these sections hold the evidence-rich detail.＋
WHAT THE BIG COMPANIES ACTUALLY BUILTSame label.
Same label.
Different machine.
These are not interchangeable “AI developers.” Each company picked a narrow loop around its own repositories, signals and review culture. The badge is the public Day‑2 boundary for that named system.
Generated monitor → Datadog alert → sandbox reproduction → fix proposal → human merge.
Fleet migration and verifier machinery, not live application incident response.
Expanding from review into CI triage and on-call tasks; no unattended diagnosis-to-remediation-to-rollback loop is demonstrated.
No disclosed production-operations loop.
Backports and CI repair are repository Day‑2, not live service operations.
THE FIELD AT A GLANCE · FIVE-SECOND EXPLANATION
Factories stop at different states.
Every row starts with intent. The marker shows the furthest state the public evidence supports for that family. “Day 2” is not a vibe: it means the loop reaches a living domain and proves the effect.
⚒Product changeIssue → code → PR / patch3PUBLIC EDGEreviewable change23 examples · 0 Day 2 · 12 bounded⟳TransformationInventory → rewrite → regression report3PUBLIC EDGEmigration artifact13 examples · 0 Day 2 · 5 bounded✹Day‑2 operatorSignal → hypothesis → bounded action5PUBLIC EDGEnarrow live loop7 examples · 0 Day 2 · 7 bounded◇Factory substrateRequest → session → handoff2PUBLIC EDGEworker / workflow17 examples · 0 Day 2 · 3 bounded
REFERENCE ROSTERWant the named cluster view?Open the compact company-to-family map.＋
THE FIELD AT A GLANCE
The big companies are building different loops.
Before you compare tools, identify the job they repeat. The cards below turn a crowded vendor list into four useful clusters and show the Day‑2 boundary of each one.
Product change factories23 assessed examples
Issue-to-PR systems turn bounded product intent into code, tests and a reviewable change. Stripe, Vercel, GitHub and many managed agents live here.
StripeStripe MinionsUberUber MinionDoorDashDoorDash Flux
0 evidenced12 bounded9 before Day 22 unknown
Compare this cluster →Transformation factories13 assessed examples
Migration and modernization systems repeat a known change family, where eligibility, rewrite rules and regression checks can be made explicit.
SpotifySpotify Honk / FleetshiftCognitionCognition DevinAmazon Web ServicesAWS Transform
0 evidenced5 bounded8 before Day 2
Compare this cluster →Day‑2 operators7 assessed examples
RCA, AIOps and rollout systems face the hardest seam: turning incomplete evidence into a safe, testable operational decision.
RampRamp Inspectkube-rca contributorsKubeRCAKevin Dubois / contributorsKubernetes AIOps Agent
0 evidenced7 bounded0 before Day 2
Compare this cluster →Factory substrates17 assessed examples
Open runtimes, durable workflows and worker planes provide pieces you can assemble—but you inherit the proof, identity and operations layer.
DoorDash OSSDoorDash Agentic OrchestratorGitHubGitHub Agentic WorkflowsCoinbaseCoinbase Mux
0 evidenced3 bounded13 before Day 21 unknown
Compare this cluster →THE SHARED DEFINITION
Day 2 means the loop closes.
A healthy worker, a green build or a merged PR is not Day 2. We use the label only when the named system reaches a living domain, acts within a boundary and has independent evidence that the outcome held.
NEXT · THE COMPARISON BENCHNow hold one job constant.
Now hold one job constant.
Compare three candidates.
The Factory Map explains the examples. Compare turns that map into a fair test: same mission, same questions, visible stopping point, Day‑2 status and human residue.