Anthropic’s Claude Code team describes layered review as agent-written PR volume surged: pre-PR verification, Claude review on every PR, automated security review, CI gates, risk-tiered human approvals, merge queue, flags, and fast rollback.
From June to September, merged PRs more than doubled while revert share fell over 40%. The team optimizes for minimum human review time without raising incidents, using CLAUDE.md and REVIEW.md, multi-agent finding verification, risk classifiers, and incident-derived evals. Low-risk docs/copy may merge without human approval when automated checks pass; high-risk paths require two humans.
The post argues human attention should focus on direction, architecture, and accountability (“you land it, you own it”) rather than line-by-line reading of agent diffs.