Let AI agents review each other instead of making people referee

A plan made with Claude Code was copied into Codex to catch anything the first model missed. Code changes were handled similarly: one model reviewed them, and another performed a second review.

Sharing the plan or giving both access to the same files was easy. The real burden began when the second model returned many suggestions, because a person still had to judge each one, check whether it came from missing information, and predict whether the first model would object.

This did not save much time; it turned the person into a and referee between . A better setup would let the challenge , check proposed changes, and agree on what belongs in the plan or code while keeping their discussion visible to the person in charge.

Key points

  • Sending the same plan to two models can increase the amount of required.
  • A person still has to check whether each suggestion is valid or based on missing information.
  • Direct discussion between could reduce the need for a human .
  • The ' discussion should remain visible rather than happening out of sight.
  • Token use, cost, and time savings need to be measured in a real .
Read original