Anthropic Launches Claude Code Review: A Fleet of AI Agents for Every Pull Request

Anthropic just shipped Code Review, a new feature for Claude Code that automatically dispatches a team of specialized AI agents to analyze every pull request as soon as it opens.

This isn't a basic linter or a single-pass AI scan. Multiple agents work in parallel — each focusing on a different class of issue — then a verification step cross-checks findings against actual code behavior to filter out false positives. Results are posted as inline comments directly on the lines where issues were found.

The feature is available now in research preview for Teams and Enterprise subscriptions.

How It Works

When a PR opens or updates, Claude Code Review:

1. Dispatches multiple agents in parallel, each scanning for a different issue type (logic errors, security vulnerabilities, edge cases, regressions)
2. Verifies candidates against actual code behavior to eliminate false positives
3. Deduplicates and ranks findings by severity
4. Posts inline comments on the specific lines where issues exist

If no issues are found, Claude posts a short confirmation. Reviews complete in 20 minutes on average and scale in cost with PR size and complexity.

Severity Levels

Each finding includes collapsible extended reasoning so developers can understand why Claude flagged the issue and how it verified the problem.

The Numbers

Anthropic has been running Code Review internally, and the results are significant:

• Substantive review comments jumped from 16% to 54% of all PRs
• Less than 1% incorrect findings (false positive rate)
• Large PRs surface ~7.5 issues on average
• 20-minute average review time

This means more than half of all PRs now get meaningful AI-generated feedback, with nearly zero noise.

What It Checks (and What It Doesn't)

By default, Code Review focuses on correctness: bugs that would break production. It does not flag formatting preferences or missing test coverage unless you configure it to.

You can expand what it checks using two files:

• CLAUDE.md — shared project instructions (also used by interactive Claude Code)
• REVIEW.md — review-only guidance for team conventions, things to always flag, and things to skip

Example REVIEW.md use cases:

• "Prefer early returns over nested conditionals"
• "Any new API route must have an integration test"
• "Don't comment on formatting in generated code under /gen/"

Setup

An organization admin enables Code Review once:

1. Go to claude.ai/admin-settings/claude-code
2. Install the Claude GitHub App
3. Select which repositories to enable
4. Choose trigger mode per repo:
   • After PR creation only — one review when PR opens
   • After every push — continuous coverage with automatic thread cleanup when you fix flagged issues

The GitHub App requests repository permissions for contents (read/write), issues (read/write), and pull requests (read/write).

Pricing

Code Review costs $15–$25 per review depending on PR size and complexity. Admins can monitor usage and spend via the analytics dashboard.

For teams doing 50 PRs/week, that's roughly $750–$1,250/week — which Anthropic positions as significantly cheaper than the engineering hours saved on manual first-pass reviews.

Self-Hosted Alternative

For teams that want to run reviews in their own CI infrastructure instead of Anthropic's managed service, Claude Code also supports GitHub Actions and GitLab CI/CD integration. This gives you full control over where code is processed.

What This Means

Code Review is the latest in Anthropic's pattern of building targeted tools on top of Claude Code's foundation:

• Claude Code (base) — CLI coding agent
• Claude Code Security (Feb 2026) — vulnerability scanning
• Code Review (Mar 2026) — multi-agent PR analysis

The strategy is clear: start with a powerful primitive, then ship focused products for specific workflows. Each one is built to slot into existing developer tools (GitHub, CI/CD) rather than requiring teams to change how they work.

For development teams, the practical question isn't whether AI code review works — it's whether the $15-25/review price point delivers ROI compared to the 30-45 minutes of senior engineer time a typical manual review takes.

Based on Anthropic's internal numbers (54% substantive findings, less than 1% false positives), the answer for most teams is yes.

Running a development team that could use automated code review? At Noqta, we help teams set up AI-powered code quality pipelines — whether using Claude Code Review, self-hosted CI/CD integration, or custom review workflows.

Set up AI code review for your team →

Related

• AI Now Reviews and Merges Your Pull Requests
• Claude Code Review: Complete Guide to AI-Powered Code Review
• Cursor vs Windsurf vs Copilot 2026
• Claude Code Pricing 2026