Anthropic's Cat Wu argues that code review has become the primary bottleneck in AI-assisted software engineering. While tools like Claude Code excel at generating pull requests, those PRs are not always production-ready. Speaking at the recent AI Codecon, Wu outlined Anthropic's two-pronged strategy to make AI code review a reliable part of their process.
First, they implemented a cultural shift: the engineer who authors a PR is now fully responsible for it end-to-end, including post-deployment bugs. This ensures that peer review responsibilities are shared fairly, and prevents junior engineers from flooding seniors with untested changes.
Second, all code undergoes review by a team of AI agents. Anthropic deliberately chose the most robust version of this system to catch not only errors in the changed code but also side effects in adjacent code—which Wu illustrated with examples of ZFS encryption and a seemingly routine authentication update that had severe unintended consequences.
As a result, AI code review has become a "loadbearing part" of Anthropic's workflow, ensuring quality and safety as AI-generated code accelerates development.