GitHub Copilot now closes its own review comments

GitHub Copilot can now mark its own code-review comments as resolved after a developer pushes a commit that addresses them. The update also gives the review system more tools for checking code and replaces the single reviewer in its Lite setting with an ensemble of AI agents.

GitHub announced the changes on September 11. Together, they move Copilot's reviewer closer to managing a continuing conversation about a pull request instead of dropping a static list of suggestions.

The open comments become the to-do list

Before the update, developers could fix an issue raised by Copilot and still have to close the associated discussion manually. During a rereview, Copilot now checks later commits and resolves a thread when it determines that the underlying feedback has been addressed.

Comments that still need attention remain open. That means the unresolved threads are intended to reflect the current state of the code rather than the complete history of everything the bot has noticed.

Copilot also writes a context-specific commit message when a developer applies one of its automatic fixes. The older generic message is replaced with a description based on the change being made.

The reviewer can run the code

GitHub says the review agent now has the full set of shell tools in the Copilot software development kit, operating behind the Copilot agent firewall. It can run builds and tests, execute targeted scripts and retrieve information from tools and APIs available to it.

That matters because many defects cannot be confirmed by reading a diff alone. A change that looks correct can fail a test, break a build or behave differently when it interacts with another part of a repository.

GitHub said its experiments produced more positive developer feedback, more high-severity findings and fewer minor complaints. The company did not publish the sample size or a full methodology in the announcement.

One review, several agents

The Lite effort level now assigns several agents to the same review. Each examines the code from a different perspective, and Copilot combines their findings into a single response.

In GitHub's experiments, the ensemble increased the average number of addressed comments per review by 47 percent for high-severity findings, 31 percent for medium-severity findings and 11 percent for low-severity findings. Review cost fell by about 8 percent.

Those figures measure comments developers addressed, not an independent defect-detection benchmark. They nevertheless show why software companies are experimenting with agent teams even for apparently simple tasks: parallel reviewers may catch different classes of problems, while a final system removes duplicates.

The features change the quality of reviews without changing how developers request them. The unresolved question is how often Copilot will close a discussion too early, a failure mode that will only become visible as teams use the rereview workflow on real repositories.