How often do you review your AI code with another AI? I've noticed Copilot catching some really solid issues in code from GPT-5.6-Sol-High, so it seems like I should be running that "rinse" cycle a few more times to keep things clean.

Hi Nikhil,

We check every commit and PR with 3 different agents run by multiple models powered by code graph-search.

In principle I agree that the new flagship models are less chatty and find the biggest issues. Worth being a little careful with the "rinse cycle" — we see different models disagree with the verdicts even from flagship models and now we offer a LLM council, models need to agree before surfacing issues.

I’ve found the rinse cycle useful too, but I think there’s an interesting point where adding more reviewers stops being the real question.

If GPT produces the change and Copilot reviews it, we still need some way to know which review findings actually mattered and which were just model disagreement.

I’m starting to think the stronger loop is: generate → review → verify against tests/evidence → record what changed because of the review.

Otherwise it’s easy to end up with several models agreeing with each other without necessarily increasing confidence in the code.