Human review is the bottleneck we keep on purpose
Every engineer here works with AI coding agents. They draft migrations, write the first version of a test suite, refactor the tedious parts and leave review comments before a person ever opens the pull request. On a good week, the agents produce more code than a squad could read carefully.
That last sentence is the whole problem.
When code gets cheap to write, the scarce resource moves to the attention of someone who understands the system well enough to say “this is right” and mean it. Most teams we meet try to protect throughput by thinning out review. We went the other way.
What we actually do
A human owns every merge. Agents can open pull requests, but only someone who owns that part of the system can merge one.
Small changes, by design. An agent will happily produce a two-thousand-line diff, and nobody reads those properly. Squads slice work into changes a person can hold in their head, and the agents work inside that shape.
Review starts from intent. The ticket says what should change and why. The reviewer checks the diff against that, then against the rest of the system. Formatting and style are the machine’s job, so people spend their attention on behavior, data and failure modes.
Tests before touching legacy code. Before an agent changes an old module, it writes characterization tests that pin down what the code does today. A human reviews those tests first. It’s the cheapest way we know to refactor something nobody fully remembers.
What it costs, and why we pay it
This does cap how fast a squad can go. A squad with agents still ships noticeably more than one without, just less than it could if everyone rubber-stamped. We think that’s the right trade. A subtle bug in a payments flow or a data pipeline costs far more than a few days saved, and our clients live with the code long after we hand it over.
It also keeps people sharp. Engineers who only approve generated code slowly lose their feel for the system. Engineers who read critically, push back and sometimes rewrite keep it.
The short version
Let agents do the typing. Keep humans accountable for what ships. Design the workflow so the human part stays possible.