Claude Code finds 7.5 issues in PRs

Anthropic on March 9 introduced Code Review for Claude Code, a new pull-request review feature that dispatches multiple AI agents to inspect code changes in parallel. The company said the system is modeled on the review process it uses internally and is available in research preview for Team and Enterprise customers. (claude.com) Anthropic said the product is designed to catch bugs that can slip past quick human reviews, especially on large or complex changesets. Anthropic said it has been running the system internally for months and now uses it on nearly every pull request. (claude.com) The company said substantive review comments appeared on 54% of pull requests after deployment, up from 16% before. Anthropic also said the tool does not approve pull requests on its own and leaves the final merge decision to human reviewers. ### How does the review system work on an actual pull request? Anthropic said Code Review starts when a pull request is opened and then sends out a team of agents to analyze the diff. The agents look for bugs in parallel, verify suspected bugs to reduce false positives, and rank findings by severity before posting a single overview comment and inline notes on the pull request. (claude.com) Reviews scale with the size of the change, with larger pull requests getting more agents and deeper analysis. Anthropic said the average review takes about 20 minutes. The company described the feature as a deeper and more expensive option than its open-source Claude Code GitHub Action, which remains available separately. (claude.com) ### What are the numbers Anthropic is using to sell it? Anthropic said 84% of large pull requests — defined as more than 1,000 lines changed — produced findings, with an average of 7.5 issues per pull request. The company said small pull requests under 50 lines produced findings 31% of the time, averaging 0.5 issues. Engineers marked fewer than 1% of findings as incorrect, Anthropic said. (claude.com) Anthropic also gave a specific internal example. The company said Code Review flagged a one-line production-service change as critical because it would have broken authentication, and the issue was fixed before merge. (claude.com) Anthropic said the engineer later said they would not have caught the problem on their own. ### Where else has Anthropic pointed to real-world use? Anthropic cited TrueNAS as an early access customer example. The company said Code Review surfaced a pre-existing bug during a ZFS encryption refactor in TrueNAS’s open-source middleware, identifying a type mismatch that was wiping the encryption key cache on every sync. (claude.com) Anthropic presented that example as evidence that the tool can find defects in adjacent code, not only in the changed lines under review. Anthropic also says Claude Code more broadly is already used heavily inside the company. On Anthropic’s product page, the company says the majority of its code is now written by Claude Code, with engineers focused on architecture, product decisions and managing multiple agents in parallel. (claude.com) ### What is verified about the newer Claude Code changes mentioned alongside this? GitHub issue and release records show one of the adjacent claims is supported. A user request to expose `duration_ms` for tool execution in Claude Code hooks was filed and then closed, and the request described the field as already measured internally but not previously exposed to hook inputs. (claude.com) GitHub release notes show Claude Code itself is still being updated frequently in May 2026, with version 2.1.143 posted on May 16. The release excerpt visible in GitHub search results does not, by itself, verify the separate claim that MCP servers now run in parallel, so that point could not be independently confirmed from the primary sources reviewed here. (anthropic.com) May 16 GitHub release notes show Anthropic continuing to ship Claude Code updates, while the company’s March 9 announcement still lists Code Review as a research preview for Team and Enterprise customers. (github.com) Anthropic’s public product and blog pages remain the main places where the company has described rollout status and performance figures for the feature. (github.com)