Anthropic Launches Multi-Agent Code Review for Claude Code
Anthropic Launches Multi-Agent Code Review for Claude Code
Anthropic has introduced a new Code Review feature for Claude Code, utilizing a multi-agent AI system to automatically analyze developers’ pull requests and catch critical bugs before human review.
Contexto
As AI coding assistants increase developer output, code reviews have become a major workflow bottleneck. Anthropic engineers themselves saw a 200% increase in code output over the past year. To address this scaling issue, they productized their internal multi-agent review methodology, making it available in beta for Claude Code users on Teams and Enterprise plans.
Detalhes
The new Code Review system runs automatically on new pull requests via a GitHub app integration. When a PR is opened, the tool spawns multiple AI agents that work in parallel to analyze the code, verify findings to filter out false positives, and rank the severity of issues. In roughly 20 minutes, the system provides a comprehensive summary comment alongside inline comments for specific problems.
During internal testing at Anthropic, the system proved highly effective:
- “Substantive” review comments increased from 16% to 54%, effectively catching nearly three times as many bugs before they could cause damage.
- For large PRs with over 1,000 changed lines, the system surfaces findings 84% of the time.
- Accuracy is exceptionally high: engineers agree with the findings over 99% of the time (less than 1% are marked incorrect).
Code Reviews are billed based on token usage, averaging between $15 and $25 per PR depending on size and complexity. Knowing that automated reviews run on every PR could quickly get expensive, Anthropic provides administrative controls such as monthly spend caps and repository-level activation to help teams manage their budgets.