AI Code Review at Agent Speed: Inside the Claude Code Team's 7-Layer Pipeline
Source: Author One pull request for the Claude Code CLI spent a whole night in a loop. An agent wrote a fix, a second agent reviewed it and found a problem, and the cycle repeated more than fifty times until someone stopped it by hand. A git blame on the reviewer's repeated findings showed that about 59% of them pointed at code that had been there since the very first pass. That story comes from the Claude Code team's October write-up on how they review code, and it shows the risk in automating review: it can turn into its own endless loop. Their results are still striking. Claude writes about 90% of Claude Code's own code, and between June and September 2026 merged pull requests more than doubled while the share that had to be reverted fell by more than 40%. Why line-by-line review stopped scaling Across Anthropic's monorepo, changes landing per day grew from about 2,500 in May 2026 to more than 30,000 in September. Reading all of that line by line is impossible. So...