Technology
Anthropic shifts Claude Code review toward automated checks and human risk decisions
Anthropic says Claude Code’s review process now moves most routine inspection to automated agents while reserving human attention for higher-risk decisions. The company reports that merged pull requests more than doubled from June to September while the share reverted fell by over 40 percent. Its system combines agent reviews, verification, risk tiers, security checks, merge-queue revalidation and staged releases. The account also documents persistent problems, including noisy recursive reviews and the absence of automatic approval tightening after incidents.
Editorial summary · Publisher article
What matters
- Every pull request receives automated bug and security review, but agent approval does not substitute for a required human approval.
- Risk classification determines review requirements, with sensitive paths and CI files requiring human review.
- Anthropic tracks reverts, incidents, comments and review rounds as safeguards against simply shifting costs downstream.
- One review loop exceeded 50 rounds, illustrating how agent authors and reviewers can fail to converge.
Context & caveats
- The performance comparisons are Anthropic’s internal measurements and are not independently audited in the supplied material.
Why this made the edition
Anthropic provides unusually detailed operational evidence on agent-heavy code review, including throughput, reverts, review layers, security checks and known failure modes; the results remain an internal company account.
Source material