On March 9, 2026, Anthropic officially launched Code Review, a sophisticated multi-agent system designed to act as a second set of eyes for software developers. Integrated directly into Claude Code, the tool is specifically engineered to solve the “review bottleneck” created by the recent explosion of AI-generated code.
The Problem: The “Vibe Coding” Surge
The rise of “vibe coding”—where developers use plain language to generate vast amounts of software—has led to a 200% increase in code output per engineer over the last year. However, this speed has overwhelmed human reviewers, leading to “skims” rather than deep reads of Pull Requests (PRs).
Code Review aims to bridge this gap by performing the deep, tedious analysis that humans often miss.
How It Works: A Team of Virtual Architects
Unlike standard linters that check for formatting or style, Anthropic’s new system uses a multi-agent architecture to simulate a high-level peer review.
- Parallel Analysis: When a PR is opened, several AI agents scan the code simultaneously from different perspectives (e.g., one checks logic flow, another hunts for edge cases, and a third verifies compliance with
CLAUDE.md). - Logic Over Style: The system ignores “bike-shedding” (debates over tabs vs. spaces) and focuses exclusively on logical and functional errors that could break production.
- Direct Feedback: Findings are posted as inline comments on GitHub, categorized by severity:
- 🔴 Red: Critical bugs.
- 🟡 Yellow: Potential concerns or smells.
- 🟣 Purple: Pre-existing historical bugs that the current PR might have touched.
Proven Performance
Anthropic has used this tool internally for months with impressive results:
- Detection Rate: In large PRs (>1,000 lines), the tool identifies issues 84% of the time.
- Accuracy: Engineers marked less than 1% of the AI’s findings as incorrect.
- Substantive Impact: Since deployment, the percentage of PRs receiving meaningful, high-signal feedback at Anthropic has jumped from 16% to 54%.
“Reviews average around 20 minutes, turning a ‘skim’ into a deep, trusted audit before a human even touches the code.” — Cat Wu, Anthropic Head of Product
Availability & Cost
The feature is currently in Research Preview for users on Claude for Teams and Enterprise plans.
| Feature | Details |
| Integration | Direct GitHub App support |
| Cost | $15–$25 per review (based on token usage) |
| Control | Monthly organization caps and repository-level toggles |