Definition
Software factory
A software factory is an engineering setup in which coding agents pull work from a queue, build it inside a harness of project rules, and hand it to verification and release gates without a human writing the code — Level 5 of the autonomy ladder. Stripe’s Minions are the public example: over 1,300 merged PRs a week, agent-written and human-reviewed (stripe.dev, Feb 2026).
Why it matters
A factory changes which half of the work is scarce. Generation stops being the bottleneck and reading starts: Faros AI’s 2026 telemetry measured median time to first review up 156.6% and pull requests merged with no review at all up 31.3%. What holds a factory up is a verification oracle the team trusts more than the model — Stripe runs its agents against over three million existing tests and still sends the branch to a human after two rounds of CI.
How each tool handles it
- Cursor
- Cloud Agents and Automations run work without an open session; the
/loopskill added in Cursor 3.5 keeps a local agent going until an outcome is reached. - Claude Code
-
claude -p "task"is the pipeline step, hooks are the checks the agent cannot skip, and/goalends a run on a condition a separate evaluator model confirms. - Codex
-
codex execis the non-interactive step and/goalruns plan, act, test and review until the condition is met; sandbox policy is configured separately from approvals.
Frequently asked questions
Is a software factory the same as a dark factory?
No. A dark factory is Dan Shapiro’s name for the corner where nobody reads the code at all — "humans are neither needed nor welcome" (The Five Levels, January 2026) — the rung Simon Willison identified StrongDM with in January 2026. Most published factories, Stripe’s included, keep a human reviewing every merge. Two axes, not one — how much agents write, and how closely people still read.
How many pull requests a week does Stripe merge this way?
Over 1,300, as of 19 February 2026 — pull requests produced end to end by its Minions agents, human-reviewed, containing no human-written code. They run against a preexisting battery of over three million tests, which is the part most pipelines copy last.
Can a small team run one?
The constraint is the oracle, not the headcount. Shapiro places the few teams at Level 5 at under five people (The Five Levels, January 2026), and what makes that possible is checks trusted more than the model. Without that, more agents only move the queue from writing to reading.