GitHub Copilot's code review now closes its own comment threads and grades harder
Comments now auto-resolve when a follow-up commit addresses them, and Copilot's lower-effort review tier switched to multiple agents voting on one review -- a change GitHub says raised how often high-severity findings actually get fixed.
GitHub said on September 11, 2026 that Copilot's pull-request code review now automatically resolves its own comment threads once a subsequent commit addresses the feedback, keeping open threads limited to what's still unresolved. The same update adds contextual, change-specific commit message suggestions in place of generic auto-filled text, and moves the "Lite" review effort level onto an ensemble of multiple agents instead of a single one. Per GitHub's own changelog, "Each agent contributes its own perspective on the code" before Copilot merges their findings into one review.
What the ensemble actually bought
GitHub's changelog attributes concrete gains to the new backend: the average number of addressed comments per review rose 47% for high-severity findings, 31% for medium, and 11% for low, while review cost fell by about 8%. The review process now also draws on a broader set of shell tools -- including running builds and tests -- to validate its own findings before surfacing them, rather than relying purely on static reading of the diff.
Why "addressed" is the number to watch
A comment getting auto-resolved isn't the same as a bug getting fixed -- it just means a later commit changed something the thread was about. GitHub's own framing ties the ensemble change to more high-severity findings actually getting acted on, not just more comments getting posted, which is the harder number to move and the one worth checking against a team's own defect data before crediting the feature. For engineering leaders already weighing Copilot's review against Cursor's BugBot or Claude Code's review tooling, addressed-comment rate by severity is a more useful yardstick than review volume alone.