Claude Code's AI-First Code Review Process

Claude Code writes roughly 90% of its own codebase, which creates a volume problem: Anthropic's monorepo grew from ~2,500 code changes per day in May to over 30,000 by September 2026. To handle that, the team rebuilt its review process from the ground up.

The core shift: stop treating 'did a human read it?' as a quality proxy, and instead ask 'which check—automated or human—is the right one to catch this specific issue?' Review now runs continuously, like spellcheck, rather than as a gate at the end of a PR. Most changes are reviewed and corrected by Claude at least once before any human sees them.

The stated goal is for most PRs to merge with at most one human in the loop—who can be the author—without increasing reverts or incidents. The team tracks human approvals per PR, human comments, and wait time on reviewers, with revert and incident rates as guardrails. Between June and September, merged PRs more than doubled while the share that had to be reverted fell by more than 40%.

The team's engineers—Addy Osmani, Morgan Westlee Lunt, and Qing Wang—frame human attention as the truly scarce resource. Saving review time only matters if it doesn't just resurface later as incident-fixing time. The post also acknowledges review's second job—building shared team context—which the automated-first approach doesn't fully replace.