What should an AI pull request reviewer do, and what should a person keep?

How it works

The Code Reviewer agent reads the diff and the linked issue, runs the checks, and leaves notes a person would recognise: a missing test, an unhandled case, a naming slip. Risky changes are marked for an engineer. The review is a record of steps, so you can see what it read and what it ran.

What a person keeps

Merging. Changes to how the system is designed. Anything the reviewer is unsure about. Over time, let the reviewer approve the small changes alone and read its record afterwards.

What to watch for

A reviewer that comments on everything is ignored within a week. Look for specific notes and a clear line between “fine” and “needs a person”.

An example

The Reliability Engineer agent opens a fix for NW-103 at 03:41. The Code Reviewer leaves two notes, one on naming and one on a missing log line, and marks the change safe. Tom merges in the morning.

Example data is invented unless a customer is named.