10/19/2026
๐ฌ๐ผ๐๐ฟ ๐๐ ๐ฐ๐ผ๐ฑ๐ฒ ๐ฟ๐ฒ๐๐ถ๐ฒ๐๐ฒ๐ฟ ๐ถ๐๐ป'๐ ๐ณ๐ถ๐ป๐ฑ๐ถ๐ป๐ด ๐ป๐ฒ๐ ๐ฏ๐๐ด๐. ๐ฑ๐ต% ๐ผ๐ณ ๐๐ต๐ฎ๐ ๐ถ๐ ๐ณ๐น๐ฎ๐ด๐ ๐ถ๐ ๐ผ๐น๐ฑ ๐ฐ๐ผ๐ฑ๐ฒ.
If you let Claude review Claude's code until it finally says "looks good", this one is about you.
Anthropic's Claude Code team had one AI agent fix a pull request while another AI agent reviewed it, round after round, until the code came back clean. It never got there. They went back and forth more than 50 times in one night, until the person who started it stepped in and stopped it.
So the team checked where all those review comments came from. About 59% of them pointed at code that had been there since the reviewer's very first look. The reviewer wasn't finding new bugs. Each round it spotted a few more old ones it had missed before, so it never ran out of things to say.
An AI author makes this worse. A person can wave off a small nit in seconds. An agent fixes every single one, which means a new commit and another round of review.
Here's what the team changed, and what we'd copy.
/1 ๐ฆ๐๐ผ๐ฝ ๐ฎ๐ณ๐๐ฒ๐ฟ ๐ฏ ๐ฟ๐ผ๐๐ป๐ฑ๐
After three rounds, their reviewer asks the author to step back and rethink the approach. If a change still fails after three passes, the plan is usually the problem, not the last line.
/2 ๐ฅ๐ฒ-๐ฟ๐ฒ๐๐ถ๐ฒ๐ ๐ผ๐ป๐น๐ ๐๐ต๐ฒ ๐ป๐ฒ๐ ๐ฐ๐ต๐ฎ๐ป๐ด๐ฒ๐
After the first few passes, the reviewer only looks at what changed since the last round. Old code stops getting a fresh batch of complaints every time.
/3 ๐๐น๐ฎ๐ด ๐๐ต๐ฒ ๐๐๐ฝ๐ฒ ๐ผ๐ณ ๐ฏ๐๐ด, ๐ป๐ผ๐ ๐ผ๐ป๐ฒ ๐น๐ถ๐ป๐ฒ
When the reviewer spots a problem like a missing null check, it names the whole kind of bug once. The author fixes every copy in one go, instead of one copy per round.
/4 ๐ช๐ฟ๐ถ๐๐ฒ ๐ฎ ๐๐ต๐ผ๐ฟ๐ ๐ฅ๐๐ฉ๐๐๐ช.๐บ๐ฑ
It tells the reviewer what failure looks like in your repo, caps the nits and lists the paths to skip. It also asks for a file and line as proof, and for a summary that says plainly when there's nothing to fix.
It worked. Pull requests that pass after one round went from 65% to 78%. The ones that needed five or more rounds dropped from 8.9% to 1.5%.
โ๏ธ More review rounds feel safer, but after three you're mostly paying to rediscover bugs that were already there.
โ๏ธ Try this today. Add one line to your review prompt that says "on re-reviews, only flag lines that changed since the last round, and say 'nothing to fix' when there's nothing."
How many rounds does your AI reviewer run before you stop it? ๐
๐ Getting AI to write and review code is the easy part. Building AI systems that keep working after the demo is the hard part, and that's what our book for AI engineers is about. Join the waitlist and we'll tell you when it's out ๐
๐ https://www.louisbouchard.ai/book/?utm_source=fb-tai&utm_medium=social&utm_campaign=book
