Code review breaks when code is produced faster than humans can read it. Anthropic’s answer is not “review harder,” but to minimize human attention without increasing rollbacks or incidents.
Its engineering workflow puts each change through seven layers of checks. The most useful ideas:
1. Evidence before submission
An agent must show that the task is complete before opening a PR, including tests, screenshots, or other relevant proof.
2. AI review with context
Every diff is checked against both the repository and the PR’s intended outcome. Multiple agents search for bugs, another group votes on the findings, and only issues judged real by the majority are surfaced. Pure style comments are discarded.
3. Approval based on risk
Low-risk changes can merge automatically, ordinary features need one human approval, and the highest-risk paths require two.
4. Safer rollout
Merge queues, feature flags, internal releases, and gradual deployment make problems easier to detect and reverse. A feature flag can disable a change without another deployment.
The strongest lesson is how to stop agent review from becoming an endless loop. One PR went through more than 50 revisions overnight; Git blame showed that 59% of repeatedly flagged code had already existed in the first revision.
Their convergence rules are simple:
• Stop after three review rounds
• If problems remain, step back and reconsider the approach
• In later rounds, inspect only newly added changes
• Never repeat feedback that has already been raised
Agent-written code and agent-led review only scale when the process is designed to converge.
https://t.co/26lu7E6KAR