Bugbot now self-improves with learned rules

2026-04-08 · Source: Cursor Blog · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning · Depth: Intermediate, quick

Summary

Cursor's Bugbot, an AI code review product, has significantly improved its bug resolution rate from 52% in July 2025 to nearly 80% today, outperforming competitors like Greptile (63.49%) and GitHub Copilot (46.69%). This improvement is now driven by a new self-improvement mechanism called "learned rules," which allows Bugbot to adapt based on real-time developer feedback from live code reviews. Previously, enhancements relied solely on offline experiments. Since launching learned rules in beta, over 110,000 repositories have enabled this feature, generating more than 44,000 unique rules. These rules are derived from signals such as developer reactions (downvotes), replies explaining issues, and human reviewer comments on merged PRs, enabling Bugbot to continuously refine its focus and accuracy.

Key takeaway

For engineering leaders evaluating AI code review solutions, Bugbot's new self-improving "learned rules" feature significantly enhances its accuracy and adaptability. Your teams can enable this feature to customize Bugbot's focus on specific issues and business contexts, potentially reducing false positives and improving code quality by leveraging real-time developer feedback.

Key insights

Bugbot now self-improves by learning from real-time developer feedback in live code reviews.

Principles

• Real-time feedback improves AI performance.
• Developer interaction refines AI code analysis.

Method

Bugbot processes developer reactions, replies, and human reviewer comments on merged PRs into candidate rules, promoting or disabling them based on accumulated signal to influence future reviews.

In practice

• Enable learned rules in Bugbot for custom issue focus.
• Utilize developer feedback to refine AI code review.

Topics

• Bugbot
• AI Code Review
• Self-Improving AI
• Learned Rules
• Resolution Rate

Best for: CTO, VP of Engineering/Data, Machine Learning Engineer, Software Engineer, AI Engineer, Director of AI/ML

Related on AIssential

• ChatGPT Co-Creator's Shocking AI Timeline — Machine self-improvement will emerge domain-specifically, not as a sudden generalist capability.
• Fragments: May 5 — AI's rapid evolution demands adaptable development practices, responsible deployment, and strategic infrastructure choices.
• Building self-improving tax agents with Codex — Self-improving agents can be built by fusing practitioner expertise with AI-driven feedback loops and structured production data.
• Before You Hit Push, Review Your Own Code — Proactive self-review, augmented by AI, elevates engineering quality and team collaboration.
• Self-Improving Agents & Knowledge Graphs: The Experimental Flywheel — Manual content significantly outperformed AI-generated content, necessitating an experimental flywheel for self-improvement.
• AI Can Generate Code in Minutes. Reviewing It Responsibly Is the Real Engineering Skill — AI accelerates code generation, but shifts the engineering bottleneck to responsible review and validation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Cursor Blog.