GitHub Copilot code review now resolves its own comments after a fix commit, and runs Lite reviews through an ensemble of agents

GitHub published a changelog entry on 11 September describing four changes to Copilot code review. Two are workflow polish, one widens what the review agent is allowed to do, and one is a rearrangement of the model that runs the default review effort.

The two workflow changes

When a reviewer addresses a Copilot comment with a follow-up commit, Copilot now resolves the comment on its own during the re-review. Before this change, the comment stayed open until someone clicked resolve by hand, even when the diff clearly answered it, which cluttered the review with stale threads. GitHub also says that when a developer applies one of Copilot's suggested code fixes, the commit message is now written by Copilot from the change itself, instead of the generic auto-fill GitHub used before.

Broader tool access for the review agent

GitHub writes that the review agent "can now access the full set of shell tools from the Copilot SDK". In practice that means the agent can run builds, run the test suite, and execute scripts inside the review environment, so a review comment can point at what actually happened when the code ran. GitHub does not spell out how the sandbox is scoped for review runs. Enterprise administrators already have controls for Copilot agent sandboxing in JetBrains as of the same weekly release, and the review agent sits in the same permission model.

Ensemble analysis is now the default Lite behaviour

The larger change is in the model that runs the review. Copilot code review has a Lite effort level, which runs on every pull request by default, and a heavier Full effort level. GitHub has replaced the single-agent analyzer that produced Lite reviews with an ensemble: several agents contribute independent perspectives, and the combined output is what gets posted.

GitHub gives four numbers for the ensemble against the previous single-agent Lite. High-severity findings that reviewers went on to address are up 47 percent. Medium-severity findings are up 31 percent. Low-severity findings are up 11 percent. Review cost, which GitHub charges customers by, is down 8 percent by GitHub's rounding. All four are GitHub's own numbers on its own benchmark, so treat them as a claim about direction rather than a measurement anyone else has reproduced. The pattern is internally consistent: the biggest gains sit on the high-severity band that reviewers care about most, and the smallest sit on the low-severity band where a small increase would be noise.

For a team already using Copilot code review on every pull request, the default tier now catches more of the things a senior reviewer would flag, and does so without a jump to the Full-effort tier. Paired with the auto-resolution change, the review loop asks less housekeeping from the human on the other side of it.

Source

GitHub Changelog, Auto-resolution and analysis updates in Copilot code review, 11 September 2026.