Don't trust your AI agent. Trust your gates.

The first day I let an agent loose on the production repository, the report it handed back was flawless: fourteen files, two migrations, tests green. I skimmed it and merged. By the next morning half the pages were returning a 500 and neither of us knew why.

The model did not fail. What failed was me reading the summary instead of looking at the check. Since then the rule is the same for the agent and for me: what does not pass the gate does not go in.

An agent does not need trust, it needs a judge

When a person opens a pull request, a social review is part of the work. Someone looks, asks, stops it. An agent writes at three in the morning with nobody beside it. The only review you can give it is the one that runs on its own and says no while you are asleep.

That is a gate: a test, a typecheck, a linter, a verifier. Its value is not in what it checks but in the fact that you cannot argue with it. It returns an exit code, not an opinion.

What this looks like day to day

In this repository every task ends in the same place. The agent writes the code and, before claiming it is done, runs the full check:

1npm run check
2# tsc --noEmit && eslint . && vitest run
3# 30 files · 192 tests
4# you exit 0 or you do not exit at all

That command does not ask how anyone feels. If it exits non-zero the work is not done, however good the explanation that comes with it. And because the agent cannot edit the test to make it pass, its only way out is to fix the code.

When the automated check cannot reach, because the problem is judgement rather than syntax, a second agent comes in with the opposite brief: find what is wrong. It does not review its own work, and that distance is exactly what makes it useful. Half the time it finds something the first one missed.

What changed

At first we defended the agent's code because we understood it. Now we defend it because it passes the same tests as ours, and when it does not, there is no conversation to have.

The agent stopped being someone to watch and became someone to work with. The difference is not trust: it is having the judge in place before it starts writing.