Your coding agent shouldn't grade its own homework

Scrolling X today, the same idea kept showing up in different repos: add a review step before agent code gets pushed, and don't let the agent that wrote the code decide it's done.

I agree, and it's the rule I ended up building around.

Why self-review doesn't work

When one model writes the code and then checks it, it tends to repeat its own blind spots. The assumptions that produced the bug are the same ones it uses to review it. Rerunning the same model with "please double-check" mostly gets you more confidence, not more coverage.

What helped in practice

1. Run deterministic checks first. Linters, tests, secret scanners. They catch the boring stuff cheaply, so any model review can spend its attention on intent and edge cases.
2. Use a different model as the reviewer. Different training, different blind spots. Disagreement is the useful signal.
3. Keep states separate. "Candidate", "verified", "reviewed by a human", and "shipped" are four different things. Collapsing them is how unreviewed code ends up on main.
4. Get the source out as files. Reading a chat transcript isn't reviewing. Download the code, open it in your editor, and look at the hand-off.

How I built this into my tool

I make Brainstorm, a desktop workspace where several AI models discuss a goal, split the work, and a different model reviews before anything counts as done. You export the source and make the final call yourself.

It's BYOK: you connect your own model accounts or API keys, and we don't sell tokens. Linux is the first platform.

If you want to look: https://brainstorm.mungwaan.cn

How do you gate agent-written code today: tests only, a second model, or a human every time?