Anthropic’s Claude Code team has shifted much of routine code verification to AI agents, while keeping human approval requirements tied to the risk of each change. The workflow reflects a practical response to agents producing a growing share of the team’s code.
Before a pull request is opened, Claude runs tests, builds and linters. A separate agent then examines the change with fresh context. Multiple Claude-powered reviewers investigate possible bugs, other agents verify their findings, and a security review can block a merge when issues remain unresolved.
Human review is not removed entirely. Documentation changes and internal features protected by feature flags may pass without human approval. Medium-risk changes require one person, high-risk work requires two, and sensitive areas such as continuous-integration workflows always need oversight. Anthropic reported that merged pull requests more than doubled between June and September while the share of changes later reverted fell by more than 40 percent. Those internal figures do not prove the process will transfer cleanly to other teams, but they show how one production codebase is separating mechanical checks from higher-stakes judgment.