People decide, agents build, checks guard main

Who it's for

Engineering leaders deciding how much to hand to AI agents.

You'll be able to

Name the five places a person must say yes, and how each one is held.

Agents can do most of the typing. That does not mean they should make the decisions.

Three jobs, kept apart

A factory splits the work into three jobs, and keeps them apart.

People decide what gets built, when work starts, and what reaches main.
Agents build. They write the tests and the code, run the checks, and report. An agent does not approve its own work.
Checks guard main. One fixed set of checks runs on every pull request, for people and agents alike.

People decide, agents build, checks guard main.

The five places a person must say yes

Nothing reaches users without passing all five.

Yes	What the person does
1. The spec	Writes it, with an agent's help, and signs it off. No code before an approved spec.
2. The plan	An agent breaks the spec into small tasks, each with its test. A person approves the list.
3. The start	A person adds a "ready" label to the task. Only then may an agent pick it up. Agents never set it.
4. The checker report	AI checkers read the change and write a report. A task closes only when the gate passes and a person has read it. A fail needs a fix, or an override the person writes down.
5. The merge	A person reads the pull request and merges it. Only then does it go live.

Notice what is not on the list. A person does not type the code or run the tests by hand. The yes is where judgment is needed, not where effort is.

A person must say yes

Step	The person	The agent	The checks
Yes 1The spec	Writes it, with an agent's help, and signs it off.	Helps write it. No code before an approved spec.	Nothing yet.
Yes 2The plan	Approves the list of tasks.	Breaks the spec into small tasks, each with its test.	Nothing yet.
Yes 3The start	Adds a “ready” label to the task.	Picks up the task only then. Never sets the label.	Nothing yet.
The build	Takes over when the agent fails three times at the same problem.	Writes the tests and the code, runs the checks, and reports.	Lint, types, tests, a secret scan and a package audit, on every pull request.
Yes 4The checker report	Reads the report. A fail needs a fix, or an override the person writes down.	AI checkers read the change and write a report.	The gate must pass before the task closes.
Yes 5The merge	Reads the pull request and merges it. Only then does it go live.	Does not approve its own work.	Guard main: the gate is a required check.

Five yeses, all said by a person. Agents build, and the checks guard main.

People own the rules too

The gate, the agents' instructions and the tool settings are control files. An agent does not edit them directly. It proposes a change in a pull request, and the named owner approves it. In CI, the gate uses the settings from main, not from the change. So an agent cannot quietly loosen its own rules.

Two more rules keep the split honest. An agent never holds a key that can publish. And when an agent fails three times at the same problem, it stops and hands over to a person with what it tried.

Checks hold the line in between

People cannot read every line an agent writes. The checks make that unnecessary. Lint, types, tests, a secret scan and a package audit run the same way every time. If a check fails, the agent fixes the cause. It does not weaken the check. By the time a person reviews, the routine questions are answered, and the review can focus on whether this is the right change.

Try it