Who it's for
Engineering leaders planning how agents fit into their team.
You'll be able to
Say who owns each step of a feature, and where the review work goes.
The engineer types less code. The engineer does not do less work.
Who does what for one feature
In a factory, three kinds of workers touch every feature: the engineer, the AI, and the automatic checks. Each step has a clear owner.
| Step | The engineer | The AI | Automatic checks |
|---|---|---|---|
| Choose the work | Decides what to build next | Nothing | Nothing |
| Describe it | Answers questions, approves the spec | Asks questions, writes the spec | Nothing |
| Plan it | Approves the plan | Writes the plan, splits it into small tasks | Nothing |
| Build it | Takes over when the AI keeps failing | Writes the code and the tests | Run on every change, show a failing one as failed |
| Review it | Reads the AI reviewers' notes, approves or rejects | Reads the change, gives advice | Tests and security scans must pass first |
| Learn | Decides which problems become new rules | Drafts the fix | Turn errors and reports into new work |
The engineer is not just watching
The engineer decides what to build, approves the written spec and the plan, and reviews every change.
The engineer does less typing. The engineer does more of four things: deciding what to build, describing it clearly, judging the result and the architecture, and improving the rules.
AI reviewers give advice. A person is responsible for every merge.
Checks decide, reviewers advise
Many teams already run good checks: a spec review, tests written first, an AI code review, a person who reviews and tests. The weak point is that they depend on people and agents following the steps. Skip one, and a failing change can still get in.
A factory changes who decides pass or fail. Fixed checks, like tests and security scans, run by themselves on every change and decide. AI reviewers only advise the person. The person still approves every change, on purpose. More freedom for the AI is earned step by step, when the numbers show it is safe.
Plan for the review load
More code from agents means more to review. Published research points the same way. DORA's 2026 report on the return on investment of AI-assisted development says speed gained in development is lost in validation unless review changes. The Cortex 2026 benchmark found that the teams whose quality did not drop already treated testing, documentation and review as measured standards.
So plan for review. When the automatic checks have already passed, the reviewer can focus on judging the result and the architecture, not on what a test already proved.
Try it
Take one feature your team shipped last month. For each step in the table, write down who actually owned it. Then estimate how many hours went to review. Watch that number as agents arrive.