GitHub's Copilot code review switches to an agent ensemble

GitHub's Copilot code review now runs multiple agents on Lite-level reviews in place of one, per GitHub's changelog. For teams running Copilot, that's the difference between a bot that flags style nits and one that catches what a human reviewer would.

The high-severity gain was the biggest: 47% more addressed than the earlier single-agent setup. Medium-severity comments were up 31%, low-severity up 11%, and review cost down about 8% over the same tests.

The review agent also picked up the full Copilot SDK toolset. It can run builds and execute tests while reviewing a pull request, instead of reading a diff alone. Copilot now auto-resolves its own comments once a pushed commit addresses them, cutting the step where a human used to dismiss them by hand.

Success here means a developer acted on the comment, not raw comment volume, which rewards noise as much as signal. That's a better test, but the changelog doesn't say how much of the gain is bugs caught rather than developers giving in to the bot. It also doesn't break down false-positive rates or say how review latency changes now that a review can trigger a full test run.