Claude Code introduces multi-agent code review in beta

Claude Code announced on March 9 a new multi-agent code review feature, Code Review, now available in beta. Modeled on the system used internally at Anthropic, it focuses on deep analysis to detect bugs that standard automated checks often miss.

Addressing the code review bottleneck

Code output per engineer at Anthropic has grown 200% over the past year, creating a bottleneck in review processes. Many pull requests (PRs) previously received only shallow reviews. Code Review addresses this by deploying multiple agents to perform parallel, in-depth analysis, offering more detailed bug detection than the existing Claude Code GitHub Action.

Internally, Anthropic observed that before Code Review, only 16% of PRs received substantial review comments. After implementation, this rose to 54%. PR approval remains a human decision; the AI does not automatically approve changes.

How the multi-agent review works

When a PR is opened, Code Review dispatches a team of agents to detect bugs in parallel. The system verifies issues to filter out false positives and ranks them by severity. Results appear as a single high-level comment on the PR, with inline comments for specific bugs when needed.

Reviews scale with the size of the PR. Large or complex changes are analyzed more deeply, while trivial updates receive a lighter pass. Average review time is around 20 minutes.

Early results and real-world impact

In internal testing, large PRs over 1,000 lines had issues detected in 84% of cases, averaging 7.5 findings per PR. Small PRs under 50 lines saw issues in 31%, averaging 0.5 findings. Fewer than 1% of flagged issues were marked incorrect.

In one example, a seemingly minor one-line change to a production service was flagged as critical by Code Review, preventing a potential authentication failure that the engineer admitted they would have missed.

During a ZFS encryption refactor in TrueNAS’s open-source middleware, Code Review identified a pre-existing type mismatch bug that silently cleared the encryption key cache—a latent issue likely overlooked by human reviewers.

Cost and administration

Code Review prioritizes depth and is more expensive than the GitHub Action, with average token costs of $15–25 per review depending on PR size and complexity. Organizations can manage usage via:

• Monthly spending caps
• Repository-level control
• Analytics dashboards tracking reviewed PRs, acceptance rates, and total costs

Getting started

Code Review is available in beta for Team and Enterprise plans. Admins can enable it in Claude Code settings and select repositories. Once enabled, reviews run automatically on new PRs without additional configuration. Full documentation is available on the Claude Code site.