Who it's for
Engineering leaders deciding how far to let AI agents go.
You'll be able to
Explain what an AI software factory is, in two minutes, without hype.
Generating code is the easy part. A factory runs the whole loop.
Not a faster typist
Most talk about AI and software is about writing code faster. That part is now cheap. The hard part is everything around it: deciding what to build, knowing whether it works, and keeping a bad change away from your users.
An AI software factory is a way of working where agents do most of the building, and a fixed set of steps, checks and human decisions makes sure what they build is right.
- People decideWhat to build. They approve the spec and the plan.
- Agents buildThe spec draft, the plan of small tasks, the code and the tests.
- The gate checksThe same fixed checks on every change.
- A person mergesReviews the change, then merges it to main.
- ProductionThe release goes live and is watched.
What is learned in production comes back as new tasks.
Every change takes the same path, in five parts:
- People decide: What to build. They approve the spec and the plan.
- Agents build: The spec draft, the plan of small tasks, the code and the tests.
- The gate checks: The same fixed checks on every change.
- A person merges: Reviews the change, then merges it to main.
- Production: The release goes live and is watched.
What is learned in production comes back as new tasks. The path starts again with people deciding what to build.
The loop
Every change goes through the same loop: an idea, a written spec, a plan of small tasks, the build, the gate, a person's review and merge, and production. What is learned in production comes back as new tasks.
People decide, agents build, checks guard main.
What it is not
- Not "no humans". People approve the spec, the plan, each task's release and every merge.
- Not one magic tool. It works with different AI tools and models. The way of working stays the same.
- Not trust in the agent's word. An agent saying "done" is a claim. The gate checks it.
Try it
Pick one change your team shipped last month. Write down who decided each step, and what checked the result before it reached users. Where nobody did, that is where to start.