This loop implements a change and then hands the diff to a reviewer from a different model family that argues against it on purpose.
The cycle is: build, cross-review with a different model, fix every blocking finding and review again, until the builder and the outside reviewer agree or the round cap is hit.
The idea is simple: an agent should not grade its own homework.
With two different models in play, neither can rubber-stamp itself, and code only lands when both genuinely sign off.
When to use
- On security, auth, or data-sensitive changes.
- When you want more rigor than a single review can give.
- To stop the agent from approving its own work.
How to use
Run the loop asking it to build, review adversarially with a different model, fix the findings, and land it only when both sides report no blocking issues.