Anthropic has officially launched Code Review, an AI-driven tool designed to keep pace with the massive volume of code generated by coding agents. The tool reads GitHub pull requests and posts findings as inline comments, allowing human reviewers to focus on riskier logic rather than syntax. With internal data showing over 30,000 changes merged daily at Anthropic, the company argues that traditional human-only review processes are breaking under the weight of agentic output.

Solving the Reviewer Bottleneck

The core problem Code Review addresses is simple: coding agents write faster than humans can read. Anthropic notes that when pull requests pile up, quick human reads often miss race conditions and security flaws. The new tool performs a close read of each change and the surrounding code to catch bugs that a PR exposes but didn't necessarily cause. Findings are ranked by severity and tagged as Important, Nit, or Pre-existing, ensuring that serious issues are actionable immediately. Most reviews complete within 6 to 10 minutes, meaning the feedback is ready by the time a human opens the PR.

Internal Metrics and Adoption

Anthropic released performance data from its own repositories to justify the tool's necessity. As of September 2026, the company saw a 40% drop in the Claude Code revert rate between June and September, even as merged pull requests doubled. The share of PRs receiving substantive comments jumped from 16% to 54%, a threefold increase. This data suggests that automated first-pass reviews significantly improve code quality metrics, allowing teams to handle higher volume without sacrificing coverage. The tool is billed on token usage through separate usage credits, with admin controls for monthly spend limits.

Integration and Limitations

Code Review integrates directly with GitHub, including Enterprise Server and Cloud with data residency, but it does not currently support GitLab, self-hosted environments, or per-user access controls. It is incompatible with organizations using zero data retention, HIPAA configurations, or customer-managed encryption keys. Setup requires an organization owner to enable the feature in Claude admin settings and install the Claude GitHub App. While reviews run on Anthropic’s infrastructure, access to the repo is read-only during the process; nothing is pushed to the branch. Users can customize what gets flagged by editing a REVIEW.md file, giving teams control over the review logic.

Key Takeaways

  • Code Review reduces the revert rate by 40% in Anthropic's internal tests while handling 30,000+ daily merges.
  • Reviews complete in 6-10 minutes, providing feedback before human reviewers open the PR.
  • The tool is billed separately via token usage credits and requires a Team or Enterprise plan for automated PR reviews.
  • GitLab and HIPAA-compliant organizations are currently unsupported.

The Bottom Line

Code Review is a necessary band-aid for the chaos agentic coding creates, but relying on AI to catch security flaws in AI-generated code feels like a house of cards waiting for a strong wind.