Review AI-written code with layers of agent and human checks

At Anthropic, most of the code in Claude Code is now written by Claude itself, and the team changed how it reviews code to match. Before a PR exists, Claude runs tests or builds to prove its work, and a second agent reads the diff with fresh eyes. On each PR, several agents look for bugs at once, and others vote on which findings are real. A separate security check and CI must also pass. Then a classifier reads the change and decides if zero, one, or two people must approve. Between June and September, the team merged over twice as many PRs, yet the revert rate dropped by more than 40 percent. They also learned to limit the back and forth: if a PR is still going around after a few reviews, the reviewer asks the author to rethink the approach instead of patching again.