Anthropic Reveals Claude Code's AI-Powered Review

Anthropic has publicly detailed the multi-layered code review process that its Claude Code team uses to handle an expanding volume of AI-generated contributions. Between June and September, merged pull requests to the Claude Code repository more than doubled, while the proportion of changes requiring a revert fell by over 40%. The company says Claude writes about 90% of Claude Code, prompting a fundamental rethink of how code is verified before production.

Shifting Verification Earlier in the Development Process

Rather than relying on traditional line-by-line reviews, the team has moved much of the verification earlier in the development process. Before opening a pull request, Claude runs tests, builds and linters, and a separate agent reviews the changes in a fresh context. This initial gate is designed to catch obvious problems before any human reviewer sees the code.

Once a pull request is opened, every pull request goes through automated checks, and Claude-powered reviewers inspect code changes for potential bugs. According to Anthropic’s Code Review documentation, each review runs on Anthropic infrastructure in four stages: several agents examine the diff and surrounding code in parallel to find issues, a verification step checks candidate findings against the code and filters out false positives, the remaining findings are deduplicated and ranked by severity, and the results post to the pull request.

Multiple AI Agents Independently Investigate Findings

Claude-powered reviewers inspect code changes for potential bugs, with multiple agents independently investigating findings before another group verifies them, and a separate security review can also block a merge if unresolved issues remain. The parallel-agent approach is intended to reduce false positives while ensuring that genuine issues are escalated.

By default, Code Review focuses on correctness, meaning bugs that would break production rather than formatting preferences or missing test coverage. Comments concerning only code style, diagnostic hints or code cleanup are filtered out. This priority structure aims to keep review feedback relevant to actual risk.

Risk-Based Human Approval and Review Cycle Limits

The central idea is to optimize code review around the risks a change introduces rather than the amount of code a human has personally read. Human approval depends on risk, with medium-risk changes requiring one person. The company has not disclosed full thresholds for each tier, but the system is designed to allocate human attention where it matters most.

Review cycles are capped at three rounds before the author is asked to reconsider the implementation. This constraint prevents extended back-and-forth exchanges and encourages contributors to rework a change if automated or human reviewers repeatedly flag issues.

Improved Pass Rates and Fewer Reverts

According to Anthropic, the share of internal pull requests passing after one review and fix round increased from 65% to 78%. The decline in post-merge reverts suggests that the layered approach is catching more problems before code reaches production.

Anthropic has made a version of this system available to external users. Code Review for Claude Code is now available to all Team and Enterprise customers. There is a lighter version for local use as well, /code-review. To enable Code Review, an administrator turns it on in Claude Code settings, installs the Claude GitHub App, and selects repositories.

Key Facts

Merged pull requests to the Claude Code repository more than doubled between June and September, while the proportion of changes requiring a revert fell by over 40%.
The share of internal pull requests passing after one review and fix round increased from 65% to 78%.
Anthropic says Claude writes about 90% of Claude Code.
Review cycles are capped at three rounds before the author is asked to reconsider the implementation.
Code Review for Claude Code is now available to all Team and Enterprise customers.

Implications for Engineering Teams

As AI-generated code becomes more common, Anthropic’s experience illustrates how engineering teams may need to rethink the balance between automated verification and human oversight. The disclosed process represents a departure from conventional review workflows, prioritizing parallel automated checks and risk-tiered approval over exhaustive line-by-line reading.

Addy Osmani, Morgan Westlee Lunt and Qing Wang described how the team’s review process has evolved alongside the growing volume of AI-generated code. Their account offers insight into how one of the largest deployments of AI-authored code manages quality and reliability at scale.

Sources

Written by

Mike Gingerich

Author

Author: Mike Gingerich, President of web firm Digital Hill, Co-Founder of TabSite .
Digital and Social Media Marketer, Speaker, and Business Consultant. Part geek, part marketer, total digital junkie! Seeking to add value, make the complex simple, and leave a positive impact.

Follow me on twitter: @mike_gingerich.