Anthropic: how Claude Code reviews code when AI writes 90% of it

claude.dev (Anthropic) · rss · 2026-10-15

Anthropic details how the Claude Code team handles code review when Claude writes 90% of its own code. Daily changes in Anthropic's monorepo jumped from 2,500 in May to 30,000+ in September, while Claude Code's merged PRs doubled and revert rates fell over 40%.

Core ideas

"Did a human read it?" is a poor proxy; the real question is what could break and which check should catch it
Review now runs continuously like spellcheck — most changes are machine-reviewed and fixed before a human looks
Human time is the scarce resource; revert and incident rates are the guardrails

The layered system

Verification before PRs: give Claude something runnable (tests, builds, screenshots); Boris Cherny says this makes results 2-3x better; subagents review diffs with fresh context; CLAUDE.md and REVIEW.md encode conventions
AI review on every PR: the public Code Review for Claude Code is GA on Team/Enterprise; substantive review comments rose from 16% to 54% of internal PRs; multiple agents hunt bugs, then a voting group filters findings
Blocking security review: PRs can't merge with open security findings; replicable via the open-source GitHub Action or /security-review
Standard CI: lint, types, multi-platform tests, builds — dozens of checks before any human
Risk-scaled human approval: docs vs. auto-updater changes get different scrutiny

A first-hand engineering retrospective with a highly replicable workflow.

Original post →

Anthropic: how Claude Code reviews code when AI writes 90% of it

More from coding & agent