The phrase “software factory” now describes company operating systems, public experiments, products, institutional programs, and open implementations. This map separates those objects so readers do not compare a vendor platform with an internal way of working as if they were the same thing.
What qualifies for the AI software factory map?
An entry needs a public first-party or institutional source describing a repeatable software-production system, a platform for the wider factory loop, or an implementation readers can inspect. A generic coding agent is a worker, not automatically a factory.
Inclusion describes the public evidence. It is not an endorsement or ranking. Product capabilities and company results stay labeled First-party report until independent evidence supports a stronger claim.
| Entry | Category | What the source says or shows | Source window |
|---|---|---|---|
| Robert W. Bemer's software-factory work | Institutional lineage | Observed artifact. A secondary historical profile recording early work to define systematic software production as a factory problem. | 2026-08-31 |
| Air Force Software Ecosystem | Institutional lineage | First-party report. An official index of mission software organizations, shared platforms, and DevSecOps software factories including Platform One and Kessel Run. | 2026-08-31 |
| Uber Software Factory | Operating case | First-party report. Uber's account of how it organizes coding agents, shared context and tools, managed jobs, quality checks, and cost per result. | 2026-08-31 |
| Ramp Inspect | Operating case | First-party report. Ramp's account of an internal background coding agent built around prepared remote environments, company context, executable feedback, and human review. | 2026-09-01 |
| Stripe Minions | Operating case | First-party report. Stripe's account of background coding agents running on standard development computers, scoped context, repeatable workflow steps, automated checks, and human review. | 2026-09-01 |
| Vercel eve | Operating case | First-party report. Vercel's account of a shared framework used for more than 100 internal production agents, with resumable runs, sandboxes, permissions, activity records, and evals. | 2026-09-01 |
| Block Builderbot | Operating case | First-party report. Block's account of a shared Goose-based workflow that turns Slack messages or tickets into tested pull requests for people to steer and review. | 2026-09-01 |
| Cloudflare AI engineering stack | Operating case | First-party report. Cloudflare's account of a company-wide route to models, identity, context, approved tools, cost records, permissions, and AI-assisted review. | 2026-09-01 |
| Spotify Honk and Fleetshift | Operating case | First-party report. Spotify's account of adding agent judgment to established systems for repository ownership, fleet-wide changes, verification, pull requests, and safe merging. | 2026-09-01 |
| Tempo Software Factory | Operating case | First-party report. Tempo's example of product and error signals entering human prioritization before agents work in parallel and return pull requests for review. | 2026-09-01 |
| Airbnb LLM migration pipeline | Operating case | First-party report. Airbnb's completed 3,500-file migration line used per-file states, executable checks, bounded retries, rich context, and a manual finish for the long tail. | 2026-09-01 |
| Amazon Q code transformation | Operating case | First-party report. Amazon's historical fleet-modernization line analyzed, planned, transformed, built, tested, and presented Java upgrades for developer review; its public product path is now changing. | 2026-09-01 |
| Anthropic AI-native development lifecycle | Operating case | First-party report. Anthropic's account connects repository guidance, isolated execution, layered agent and deterministic review, risk-based human gates, staging checks, and incident response. | 2026-09-01 |
| Cognition Devin | Operating case | First-party report. Cognition's account of Devin building Devin covers Slack and issue intake, review on every pull request, bounded repair, recurring audits, incident investigation, and playbooks. | 2026-09-01 |
| Cursor cloud and security agents | Operating case | First-party report. Cursor's account combines isolated parallel implementation with visual proof artifacts and event-driven review and security lines. | 2026-09-01 |
| Dropbox Nova | Operating case | First-party report. Dropbox's shared platform separates model judgment from deterministic test, retry, branch, and publication control across interactive and background work. | 2026-09-01 |
| GitHub Copilot coding agent | Operating case | First-party report. GitHub's account shows assigned issues becoming draft pull requests inside the core github.com repository under ordinary review and repository rules. | 2026-09-01 |
| Linear Coding Sessions | Operating case | First-party report. Linear uses the issue tracker as the dispatch and review spine for managed coding sessions that return diffs, screenshots, recordings, and pull requests. | 2026-09-01 |
| LinkedIn CAPT | Operating case | First-party report. LinkedIn's CAPT turns internal tools and institutional knowledge into reusable executable playbooks for more than 1,000 engineers. | 2026-09-01 |
| Meta Capacity Efficiency agents | Operating case | First-party report. Meta's specialized line connects production performance signals, shared tools, encoded expert skills, candidate fixes, validation criteria, and engineer review. | 2026-09-01 |
| OpenAI Codex and Symphony | Operating case | First-party report. OpenAI's account combines an agent-legible, mechanically constrained repository with project-board orchestration, isolated workspaces, retries, and human review. | 2026-09-01 |
| Replit internal agent fleets | Operating case | First-party report. Replit's account of manager agents connects secure execution, company systems, parallel verifiable loops, risk-rated review, incidents, and learning from feedback. | 2026-09-01 |
| Shopify World, River, and Aquifer | Operating case | First-party report. Shopify's account joins a reproducible monorepo, public-to-the-company Slack work, durable agent sessions, disposable sandboxes, shared profiles, and specialist factory lines. | 2026-09-01 |
| StrongDM Software Factory | Operating case | First-party report. StrongDM's account of starting from a clear specification, testing realistic scenarios, simulating dependencies, and letting agents carry work end to end. | 2026-08-31 |
| Ona Software Factory | Public experiment | First-party report. A public experiment and livestream series about specifications, feedback, taste, steering, and factory operation. | 2026-08-31 |
| Augment Cosmos | Factory platform | First-party report. A product that coordinates event-triggered coding agents, human checkpoints, delivery, monitoring, and learning across repeated jobs. | 2026-08-31 |
| Warp Factories | Factory platform | First-party report. A product for defining factory intake, agent roles, permissions, evals, measurement, and improvement as versioned configuration. | 2026-08-31 |
| OpenAI Symphony | Public implementation | Observed artifact. An inspectable engineering preview that turns project tasks into isolated agent runs, records proof, and protects how changes land. | 2026-08-31 |
What do the categories mean?
- Operating case: an organization describes how it runs a factory in practice.
- Public experiment: a team exposes an evolving method, inquiry, or live trial.
- Institutional lineage: a source shows how the idea was used before today's coding agents.
- Factory platform: a product provides machinery for running the wider loop.
- Public implementation: enough working code or architecture is visible to inspect the actual boundary.
These are SWFT's editorial categories, not a market standard. An entry may move when stronger evidence changes what readers can inspect.
What does the map leave out?
Private factories with no public account are absent. Model providers, IDEs, coding agents, CI services, and developer portals are excluded unless their own material establishes a factory-level role. The map also omits market size, revenue, customer counts, and performance rankings unless a public method makes those numbers interpretable.
Absence from the map means “not established by the current public method,” not “does not exist.”
How do I suggest a correction?
Use the correction process. A useful submission names the entry, the field that should change, and the strongest public source.