On March 5 this year, GitHub announced that Copilot's code review feature has officially switched to an agent architecture.
What was wrong before
The old code review essentially looked at diffs line by line and generated generic comments. Many developers complained it was "like an intern reading code" — only looking at the surface without understanding context.
How the new version works
The new agent actively explores the entire repository: it reads related files, tracks cross-file dependencies, builds a more complete context, and then delivers review comments.
Key data points:
- 71% of reviews provide actionable feedback
- The remaining 29% choose to stay silent rather than force out meaningless comments
- Average of about 5.1 comments per review
- Focus is on correctness and architectural integrity, not code style nitpicking
From last April to this March, a cumulative total of 60 million code reviews were completed, a 10x increase.
Closed-loop auto-fix
The most clever design: issues found during review can be directly handed off to the coding agent for fixing. Copilot finds a bug → generates a fix PR → human reviews and merges. Routine small fixes can be automated in a loop, with the developer only needing to sign off at the final step.
However, note that agent code review runs on GitHub Actions infrastructure. If your organization has disabled GitHub-hosted runners, you will need to configure self-hosted runners.
Available to all Copilot paid users (Pro/Pro+/Business/Enterprise).
Sources: CocoLoop, official GitHub Changelog