Cursor says its code review agent now learns from PR activity and sees 78% of flagged issues resolved before merge

In an April 8 X post, Cursor said its code review agent can now "learn from activity on PRs to self-improve in real time." The company paired that claim with a headline metric: 78% of the issues found by the agent are resolved by the time the pull request is merged. Even as a short post, that wording is notable because it describes code review as a feedback loop, not just a one-shot model pass over a diff.

If taken literally, Cursor is saying merged PR outcomes are being used as training or tuning signals for the reviewer. That matters because most AI code review features are presented as static assistants that score a diff once and stop. A reviewer that adapts to comment threads, accepted fixes, and merge outcomes could improve precision over time, especially inside organizations with repeat patterns of bugs and review preferences. The metric Cursor chose is also more practical than raw alert counts. Counting how many comments a model leaves says little about usefulness, but reporting how many of its flagged issues are fixed before merge points much closer to developer trust and workflow integration.

The claim is interesting, but the methodology is still opaque

The post does not explain how Cursor defines an "issue found," what counts as a resolution, how large the sample is, or whether the 78% figure covers all repositories or a narrower cohort. It also does not spell out what parts of PR activity feed the learning loop or what safeguards exist against reinforcing reviewer bias. That means the announcement is best read as a strong product signal rather than an independently comparable benchmark. Even so, the direction is clear: Cursor is trying to turn AI code review from a passive suggestion engine into a reviewer that continuously adapts to real development outcomes.