Anthropic’s Claude Code team shifts code review from gatekeeping to continuous AI-assisted checks

Anthropic’s Claude Code team, including Addy Osmani, Morgan Westlee Lunt, and Qing Wang, describes a rethinking of code review that replaces the traditional end-of-the-line gate with continuous, AI-assisted checks. In a blog post, they argue that asking whether a human read the code is an imprecise way to determine what could go wrong with a change or which checks—automated or human—are best suited to catch potential issues.

The team reports that this shift has changed when review happens. Previously, code review occurred after a pull request was opened and a colleague was available. Now, review runs continuously, more like spellcheck while writing. According to the authors, most changes are reviewed and fixed by Claude at least once before a human is asked to look. This approach catches routine mistakes early and allows engineers to focus their attention on higher-level questions and deeper scrutiny.

The post is quoted and contextualized by Simon Willison on his weblog, which highlights the integration of AI agents into software development workflows. Willison’s newsletter emphasizes the operational implications of treating AI agents as having the same access as humans, while also noting the need to monitor and constrain their behavior in trusted environments.