Guides / introductory
What a software factory is
The agent-native definition, the older DevSecOps meaning, and the loop that actually has to close.
The phrase is not new. Defense and enterprise DevSecOps programs have called a standardized, paved CI/CD path a “software factory” for years. That meaning is still in use. It is not what this site is about.
In 2026 the term was pulled toward a different object: a system that takes a signal — an issue, a spec, a customer report, a failing check — and produces deployed, verified software, with AI coding agents doing the execution and engineers defining intent and reviewing results.
If the output is a chat transcript, you do not have a factory. If the output is a pull request that nobody can run, you do not have a factory. If an agent can merge because the tests were green in a context the team cannot reproduce, you have a liability.
The loop
A factory is a loop, not a tool.
- Intent. A human or an upstream system states what should change and what “done” means.
- Orchestration. Work is assigned, sequenced, retried, and handed between agents or steps.
- Execution. An agent reads the repository and produces a diff.
- Sandboxing. That work happens on a computer the team can inspect, not on a laptop that also holds production credentials.
- Verification. Tests, hooks, review, and merge policy decide whether the diff is software.
- Observability. Someone can answer what ran, what it cost, and whether the factory is getting better.
Most products in this directory occupy one or two layers well and market the rest. That is fine. A factory is usually assembled. The mistake is believing a single vendor page is the assembly.
What this is not
It is not “developers type faster.” Autocomplete is a tool. A factory is an operating model.
It is not unsupervised autonomy as a moral good. The interesting design question is where humans sit: specifying, steering, reviewing, merging. “A human merges every pull request” is a factory policy. “The agent pushes to main” is a different policy. Write it down.
It is not a synonym for one company’s product. Factory.ai named the category after itself. Cursor, Warp, Claude Code, Trevize, and a long list of harnesses and sandboxes all sit on the same stack. This site exists so those claims can be compared without a sales deck.
How to start
Do not begin with a platform bake-off. Begin with one signal you already trust — a well-typed bug, a failing CI job, an issue template your team actually fills in — and force the loop to close on that signal. If you cannot name the pass condition, you are not ready for an agent. If you cannot name the computer the agent will use, you are not ready for parallelism. If you cannot name who merges, you are not ready for production.
The methodology page explains how we assess the tools that show up in that loop. The directory is the catalog. The layer pages are the map.