Anthropic code review retained commit-bound human approval for high-risk changes

A review does not approve a label such as “safe,” detached from the code that was actually inspected. Anthropic describes an internal process in which AI reviewers examine code for bugs and security issues, while human signoff varies by risk. The high-risk tier calls for two human approvals. If the branch changes after review, the approved commit identity must still match the code presented for merge.

The operational evidence also describes repeated exchanges among review agents and developers, including discussion of existing code unrelated to the change. That friction matters: more model reviewers can introduce noise and reviewer burden alongside useful findings. A reported decrease in rollbacks, where mentioned, cannot establish that the review arrangement alone caused fewer defects because organization size, tooling and code selection also changed.

This is a disclosed governance design, not a guarantee that every merged change is safe. Its measurable claim concerns whether approval gates are bound to exact revisions and whether risk-tier routing is applied before merge. Independent study would need accepted vulnerabilities, false positives, review minutes and postmerge defects under matched conditions, rather than counting how many AI review comments appeared.