AI Agent News/stories/anthropic-details-how-claude-code-s-team-reviews-ai-written-code.md

Frameworks

Anthropic details how Claude Code's team reviews AI-written code

Anthropic has described how the Claude Code team reviews code as AI agents write a growing share of its codebase. The process layers automated checks, AI reviewers and risk-based human approval to cut manual review time without more regressions.

Source: www.infoq.com/news/2026/11/claude-code-review/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering

Addy Osmani, Morgan Westlee Lunt and Qing Wang describe how the team's review process has changed as the volume of AI-generated code has grown. Merged pull requests to the Claude Code repository more than doubled between June and September, while the share of changes needing a revert fell by over 40%. Most verification now happens before a pull request opens: Claude runs tests, builds and linters, and a separate agent reviews the changes in a fresh context.

Every pull request then passes several automated checks. Claude-based reviewers look for potential bugs, and multiple agents investigate each finding before another group verifies it. A separate security review can block a merge while unresolved issues remain. Human approval depends on risk: low-risk changes may need none, medium-risk changes need one approver, high-risk changes need two, and changes to sensitive areas such as CI workflows always need human oversight.

To limit noise, review cycles are capped at three rounds before the author is asked to reconsider the implementation, and later reviews cover only new changes. Anthropic says the share of internal pull requests passing after one review and fix round rose from 65% to 78%, while those needing five or more rounds fell from 8.9% to 1.5%. Incidents and reverts feed an evaluation dataset used to improve the automated reviewers.

Why it matters if you build agents

If you build coding agents, the takeaway is that checks can run mostly before a pull request opens, with human review scaled to risk rather than to the size of a diff. Capping review rounds and reviewing only new changes on later passes are patterns you can copy to keep automated feedback from piling up. These figures come from Anthropic's own repository, so they describe one team's results rather than a general benchmark.

Get stories like this every weekday

One short email with the day's agent tooling news.

Subscribe