Agent code review has stopped being a one-shot opinion. Since September 18, Copilot code review remembers findings across iterations and rules on their fate.
The GA refresh turns the overview comment from a list into state. Findings now live in three groups: "Open" — unaddressed, flagged as new when introduced by a fresh commit; "Resolved since last review" — verified fixes; "Previously missed" — issues discovered late in changes that were already there. Every finding carries a severity level, a concise title, and a link to its inline comment. Auto-resolution got smarter too: the agent honors an author's reply asking to keep an issue open, and states a reason when it closes one — "Won't Fix" or "Incorrect". Accept a batch of suggestions, and Copilot writes a meaningful commit message for you, even when the batch includes non-Copilot comments.
Tracking the fate of review comments across iterations used to be the human reviewer's manual bookkeeping: remember what was raised, verify what got fixed. Now the agent keeps that ledger — and the ledger has a line that never existed before: a reviewer declaring its own finding wrong. When an "Incorrect" verdict is right, you get less noise. When it isn't, a defect lands in the main branch while the review log stays spotless.
"Previously missed" is the most honest category of the three: same diff, another pass, new findings. It is an open admission that agent review is non-deterministic. Pretending otherwise would be cargo cult; the right move is not to flinch but to design the process around that property.
The takeaway for a team lead: two short agreements instead of blind faith in the ledger. First, write down which severities an author may close without a second reviewer — and which require a human signature. Second, once per sprint, sample the auto-resolves closed as "Incorrect" and check whether the agent was actually wrong, or wrong twice. And read "Previously missed" as a map of blind spots: whatever the agent only catches on a second pass belongs in a linter or a test. The agent now owns the state of the review; the state of the code is still yours.
📎 github.blog/changelog/2026-09-18-copilot-code-review-an-improved-review-experience · github.blog/changelog/label/copilot