Cursor Bugbot: Self-Improving Code Reviews with Learned Rules
CursorCursor's Bugbot code reviewer now learns from real pull request feedback to continuously improve its review quality through automatically generated Learned Rules. By analyzing developer reactions, replies, and human reviewer comments across hundreds of thousands of daily PR reviews, Bugbot promotes high-signal rules and disables low-value ones. Over 110,000 repositories adopted learned rules in beta, generating more than 44,000 custom rules. The system's resolution rate has climbed to 78%, up from 52% at launch.
Sources & Mentions
5 external resources covering this update
Bugbot Now Learns From Your Team's Code Reviews
Cursor has introduced Learned Rules for Bugbot, its AI-powered code review tool, marking a shift from static rule-based approaches to a dynamic, self-improving system. Rather than relying solely on manually configured guidelines, Bugbot now analyzes feedback signals from live pull request reviews to generate and refine its own review rules automatically.
How Learned Rules Work
At the core of this feature is a feedback loop: Bugbot monitors how developers respond to its comments — including reactions like downvotes on incorrect findings, replies that correct or clarify the AI's assessment, and comments from human reviewers that surface issues Bugbot missed. From these signals, the system generates candidate rules specific to a repository's codebase and team standards.
Candidate rules are promoted to active status when they prove consistently useful over time. Conversely, rules that generate negative feedback are automatically disabled — or can be manually adjusted by users. This means Bugbot becomes more accurate the longer a team uses it, without requiring developers to hand-write every custom rule.
Scale and Impact
The results so far are substantial. More than 110,000 repositories adopted learned rules during the beta period, resulting in over 44,000 distinct custom rules generated across those codebases. Bugbot's overall bug resolution rate increased to 78% — up from 52% when the product first launched.
Bugbot already reviews more than two million pull requests per month for customers including Rippling, Discord, Samsara, Airtable, and Sierra AI, making the scale of this learning loop particularly significant.
Additional Improvements in This Release
Alongside learned rules, this update ships several improvements to the Bugbot experience:
- A "Fix All" action lets users apply multiple Bugbot-suggested fixes simultaneously, reducing friction when acting on batch findings
- The settings interface has been redesigned with clearer separation between personal and team configuration
- Autofix now runs only when findings actually warrant fixes, reducing noise
- Prompting has been refined to use only relevant rules, further cutting false positives
- CI check reliability on pull requests has been improved
- Progress messages shown in GitHub PRs are now simpler and easier to read