Anthropic describes its code review system

Anthropic’s Claude Code team says most of its code review now happens before a person looks at a change. Claude checks work while it’s being written, reviews pull requests against the wider codebase, and checks whether findings are real before posting them. Automated tests, security checks and other rules can block a change from merging. People still review, with more attention going to changes the team sees as higher risk.

The team’s stated aim is to reduce the human time spent reviewing without increasing reverts or incidents. It tracks approvals, comments and how long pull requests wait for a person, while using reverts and incidents as guardrails. The post says that from June to September, merged pull requests to Claude Code’s codebase more than doubled, while the share that had to be reverted fell by more than 40 percent. Those are results from Anthropic’s own process, not a guarantee that the same setup will work for every company. The original write-up explains the layers in more detail.

For an operator, the useful point is that generated code isn’t self-validating. More code can mean more review work, unless checks catch routine problems and people know where their judgment matters most. The Hacker News discussion, with 445 points and 438 comments, suggests this is a timely question for teams adopting coding tools. This is the kind of thing we build into an AI command centre, where tools, checks and human decisions need to fit together.