What is a software factory?
A software factory is a pipeline in which coding agents pull work from a queue, build and verify it, and open pull requests with little or no human reading the diff. It is the top of the autonomy ladder, Level 5, reached one loop at a time: a loop may run unattended only when an oracle the agent cannot edit decides it is done.
Origin This site uses the name for Level 5 of the five levels Dan Shapiro set out on 23 January 2026 in “The Five Levels: from Spicy Autocomplete to the Dark Factory” (opens in a new tab), where he calls that level the dark software factory.
- In Polish
- Fabryka oprogramowania
- Closest ladder level
- Level 5, free guide
- Also searched as
- software factories
$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.
Why it matters.
A factory changes which half of the work is scarce: generation stops being the bottleneck and reading starts. Faros AI’s April 2026 report 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 an oracle the team trusts more than the model: Stripe’s Minions run against over three million existing tests and get at most two rounds of CI before the branch goes back to a person.
Software factory vs vibe coding
Vibe coding and Level 5 both remove the human who reads the diff; only Level 5 replaces that reader with verification a machine can check.
| Question | Vibe coding | Software factory |
|---|---|---|
| What it names | A way of working: accept what the model writes without reading the diff | A pipeline at Level 5, the top of the autonomy ladder: agents pull work from a queue, build, verify and open pull requests |
| Who writes the code | The model | Agents |
| Who checks it | Nobody reads the diff; the human watches the behaviour | No human reads the diff either (StrongDM: “Code must not be reviewed by humans”); machine-checkable scenarios and a satisfaction score do |
| What stops the loop | The app appears to work | An oracle the agent cannot edit decides it is done |
| Where the name came from | Andrej Karpathy, post on X, 2 February 2025 | Shapiro’s Level 5 (“the dark software factory”), 23 January 2026 |
| Where it breaks | Beyond throwaway work, nothing replaces the reader | Without the oracle, more agents only move the queue from writing to reading |
The six stations
Each station owns one decision. A station that emits no evidence is one you supervise by reading.
| Station | Decision it owns | Evidence it must emit | Symptom when it is missing |
|---|---|---|---|
| Intent | What the agent knows before it starts | A spec with acceptance criteria and a named stop condition | Plausible code that solves the wrong problem |
| Harness | What the agent may run and reach | A permission and sandbox policy under version control | A run that works because it had more access than it should |
| Loop | When the agent stops | A machine-checkable stop condition and a retry cap | Runs that never finish or finish by declaring victory |
| Graph | Who does what when there are several agents | A written topology: planner, workers, judge, isolation | Agents overwriting each other’s files or context |
| Verification | What counts as proof | Deterministic checks the run cannot edit, plus review findings | Green checks on tests the agent rewrote |
| Release | How work leaves the building | An evidence bundle on the pull request and a named merger | A merge nobody can explain a week later |
How each tool does it.
Cursor, Claude Code and Codex each name this their own way. Each note carries the date it was checked.
Cursor
Automations start Cloud Agents, which run in isolated VMs, on a schedule or from source-control, webhook, Linear, Sentry or PagerDuty triggers.
The Cloud Agents API sends a statusChange webhook when a run finishes, so your systems learn the result without polling. /loop is a bundled skill.
Level 5 guide (Cursor docs), 2026-08-28
Claude Code
claude -p runs the loop headless, hooks and Sandboxed Bash hold the harness, /goal sets the stop condition, and Routines (research preview) schedule runs.
A routine runs the same prompt as a cloud session, on a schedule, from an API call or from a GitHub event. Routines have no permission-mode picker and no approval prompts during a run, so the allowlist lives in the environment.
Level 5 guide, Claude Code 2.1.283, 2026-09-26
Codex
codex exec is the headless step and already runs with approval never; /goal is stable and on by default. Automations schedule recurring runs (OpenAI documentation, 2026-08-28).
The workspace-write sandbox restricts network access unless you enable it, and a dependency loop needs the registry, so the guide’s example adds -c sandbox_workspace_write.network_access=true. openai/codex-action@v1 wraps the same run in CI.
Level 5 guide, Codex 0.157.1, 2026-09-26
Questions about the software factory.
Is a software factory the same as a dark factory?
No: the dark factory is one end of it. Dan Shapiro’s “dark software factory” is Level 5, where “humans are neither needed nor welcome” (The Five Levels, 23 January 2026), and Simon Willison identified StrongDM’s AI division with it on 28 January 2026. Stripe’s Minions are a factory that keeps a human reviewing every merge, and this site rates that loop Level 3. Two axes, not one: how much the agents write, and what a human still reads.
How many pull requests a week does Stripe merge this way?
Over 1,300, as Stripe reported on 19 February 2026: pull requests that are “completely minion-produced, human-reviewed, but containing no human-written code”. The agents run against over three million existing tests and get at most two rounds of CI before the branch goes back to a person. That proof layer predates the agents.
Which loops may run without a human watching?
A loop may run unattended only when a named oracle decides “done”, the agent cannot edit that oracle, a wrong answer surfaces in minutes rather than weeks, and the blast radius is small and reversible. Patch and minor dependency updates and lint codemods can pass; auth, money, schema and migrations stay with a human reading code.
Can a small team run one?
Shapiro puts only “a handful of people” at Level 5, in teams of fewer than five (23 January 2026). The constraint is the oracle, not the headcount: what makes nobody-reads survivable is the strength of the verification around it. Without that, more agents only move the queue from writing to reading.
The vocabulary of AI-driven development
Every name below is defined against vibe coding: who writes the code, who checks it, and what stops the loop.
Read the long-form guide: agentic engineering vs vibe coding
Sources.
The primary sources outside this site that this page relies on.
- The Five Levels: from Spicy Autocomplete to the Dark Factory (opens in a new tab) Dan Shapiro, Dan Shapiro’s Blog
- The Five Levels: from Spicy Autocomplete to the Dark Factory (opens in a new tab) Simon Willison
- Minions: Stripe’s one-shot, end-to-end coding agents—Part 2 (opens in a new tab) Alistair Gray, Stripe
- Ten takeaways from the AI Engineering Report 2026: The Acceleration Whiplash (opens in a new tab) Faros Research, Faros AI
Keep reading.
The guides that go deeper, and the terms and comparisons next to this one.
In the docs
Related terms
Does this page answer your questions?
One tap, no sign-in.
What are you here for?
What is holding you back?
How does it read?
Thanks. If you left an email, Michał will reply.
Answering your questions?
One tap each. Close whenever you like.
no sign-in
Read the guides in the same words.
Open every guide with the Toolkit. Each term is defined once here, and the guides use it the same way.
$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.
Michał replies, not a bot. The address is used for this reply only.