Bugbot — Cursor's built-in pull request reviewer — has picked up a handful of genuinely useful upgrades. The headline is Autofix: when Bugbot flags a bug in your PR, it can now spawn a Cloud Agent (an AI that runs in the background on Cursor's servers) to write the fix, push it to a branch, and hand it back to you. No copy-pasting suggestions. No switching context. Just a fixed branch waiting for your approval.
That is the part most people will notice first. But the less obvious additions are just as useful.
You can now tell it how hard to think
Bugbot has three Effort Levels. Default is fast and cheap — good for catching obvious mistakes without burning through your usage quota. High is slower and more thorough, the kind of pass you want before merging something important. Custom lets you write your own trigger condition: something like "go deep when the PR touches authentication or payments." You pick the level per review, or set a default for the whole team.
The pricing model shifted in May 2026 from per-seat to usage-based, which means Effort Levels now have a direct cost implication. Running High effort on every PR would add up. Running it only where it matters is the smarter play.
It gets smarter the more your team uses it
Bugbot can now learn from your team's PR activity — more like building a memory of what your codebase cares about than following a static checklist. You can also teach it directly by commenting @cursor remember [fact] on any PR, and that fact becomes part of how it reviews future diffs. Team Rules, repository rules, and a project-level .cursor/BUGBOT.md file all feed into it, with a clear hierarchy so conflicts resolve predictably.
If you are thinking "that sounds like a lot of config to set up" — it is optional. Out of the box it just reviews. The rules layer is there when you want consistency across a team.
Review locally, skip the duplicate on GitHub
There is a /review-bugbot skill inside Cursor that lets you run a Bugbot review before you even push. If you then open a PR with the same diff, Bugbot on GitHub recognizes it and skips the remote review. One pass, not two. For solo developers who want the feedback earlier in the process, this is probably the most immediately practical change.
Bugbot also reads existing PR comments before it posts, so it will not repeat something another reviewer already caught. Small thing. Appreciated.