AI code review should reduce noise, not add more comments. That is the line I care about. GitHub recently moved Copilot code review onto an agentic architecture, which means it can gather more repository context instead of only staring at the changed lines. That is the right idea, but only if the output is signal.
Most teams do not need more review comments. They need better ones. A noisy review process trains developers to skim, ignore, and eventually distrust feedback. If AI review turns into a machine for generating plausible but low-value comments, it will make reviews worse rather than better.
The useful version catches the things people miss because they are tired, rushed, or too close to the change. Broken assumptions. Missing edge cases. Risky interactions. Code that does not fit the existing architecture. GitHub says the newer Copilot review flow can pull in broader repository context, and that matters because isolated review is often weak review.
I like AI code review as a second pass. I do not like it as a replacement for ownership. The developer still owns the change, the team still owns the standards, and a human still needs to think. The win is not more comments. The win is fewer surprises.