AI Code Review

A developer opens a pull request that changes how passwords are reset. Within a minute, an AI reviewer comments that the reset token never expires, that one error path logs the user’s email, and that no test covers the new branch. A human reviewer still reads the change, but starts from those notes instead of a blank diff.

Most tools work the same way: send the diff and some surrounding code to a model with review instructions, then post the findings as comments. Results depend on how much context the reviewer gets (the rest of the repository, the ticket, the team’s conventions), so many tools also read agent instruction files. False positives are common, and a reviewer that sees only the diff can’t judge whether the change is the right thing to build.

In episode 40 at 38:07, Graphite’s Greg Foster describes the basic version: feed the diff to a model and have it call out mistakes and vulnerabilities inline. He frames it as a fuzzy variant of CI (38:59): useful validation, but not a reason to accept a change on its own. Graphite builds an AI code reviewer (17:04), so this is a vendor’s view. The episode was sponsored by Galileo, which Conor plugs at 2:05.

In episode 74 at 5:59, Arize’s Laurie Voss estimates that the best AI code review tools he has seen have about an 80% hit rate: very good, but not enough to ship on, so senior reviewers still own the last 20%. That figure is his impression, not a benchmark. Episode 74 credits Inngest, Svix, Walrus Memory and G2i as the season’s sponsors at 4:38. Cisco’s Jeetu Patel argues in episode 67 at 37:48 that when agents write unlimited code, review becomes the bottleneck, and automating the first line of review moves the bottleneck to product judgment. Episode 67’s Svix read is at 24:06.

AI review is one layer. Keep tests, static analysis and a named human approver for changes that matter; see software factory for pipelines that try to remove human review altogether.