Review Agent

A review agent is an AI program that looks over a developer's code changes before a human does, flagging parts that might be problematic. These code changes usually come in as pull requests, which is the process of having a colleague review new code before it's merged into an existing project.

Think of it like an editor who reads a manuscript first and marks it up in red, but the editor-in-chief still makes the final call on whether the book gets published. A review agent works the same way: it comments on changed lines and tags how severe each issue is, but a human decides whether to accept the code. It never makes approve-or-block decisions on its own.

A review agent also stops at finding problems. It doesn't rewrite the errors it finds — turning those comments into actual fixes is still the developer's job.

As the article explains, "review is handled by several review agents running on Anthropic's infrastructure, which analyze the changes against the entire codebase, tag them by severity, and leave them as inline comments." It's easy to assume a review agent approves code or fixes it automatically, but in reality it only leaves comments — the final decision and the actual fix still have to be done by a human.

If you're using Claude Code, you can type the following in your terminal to review the comments a review agent just left.

/code-review

If you want those comments applied directly to your code, run it with the option attached.

/code-review --fix