Anthropic has launched Claude Code Review, a code review system using multiple AI agents. The starting observation: developers using Claude Code raised their productivity by 200% over the past year. But that acceleration created a problem: nobody reads the code in depth any more.
Merge requests get handled superficially with quick approvals, producing bugs in production, silent security holes and undetected regressions.
How it works
Claude Code Review goes beyond a simple syntax checker. It is a team of AI agents examining each PR in parallel. Each agent analyses the code from a distinct angle. The agents check one another to eliminate false positives. Every issue is categorised by severity.
Measured results
- Before: 16% of PRs received substantive comments
- After: 54%
- On large PRs (over 1,000 lines): 84% contain findings
- False positive rate: under 1%
A concrete example: a one-line change in production, normally waved through, was flagged CRITICAL because it broke the service's authentication.
Implications
AI Nursery, the AI development arm of AS3P, uses Claude Code as its main tool. Code Review closes the quality loop:
- Write the code: AI
- Review the code: AI, multi-agent
- Approve the merge: human
AI extends human oversight rather than replacing it. Developers write faster while verification improves: acceleration with no trade-off on quality.
Source: Anthropic.