Bugbot Learns From Live Code Reviews

Cursor's AI code reviewer, Bugbot, has significantly boosted its effectiveness by enabling self-improvement through learned rules derived from live code reviews. Launched out of beta in July 2025, Bugbot's bug resolution rate has jumped from 52% to nearly 80%, a 15-point lead over its closest competitors, according to data from Cursor Blog.

StartupHub data

Companies working on this

Profiles of the companies named in this story, with funding and a one-liner from our database.

CodeRabbit

$550

AI-driven platform that revolutionizes code reviews for enhanced quality and faster software delivery.

Greptile

$7M

AI-powered code search and understanding for developers.

Previously, Bugbot improvements relied solely on offline testing. This new system allows Bugbot to learn directly from hundreds of thousands of daily pull request (PR) reviews.

The AI now transforms real-time feedback into custom rules. These rules help Bugbot focus on specific codebase issues and business priorities, making its suggestions more relevant.

Learned Rules Drive Accuracy

Bugbot's enhanced resolution rate, now nearing 80%, showcases its improved performance. This marks a substantial leap from its initial 52% rate, positioning it ahead of other AI code review tools like Greptile and CodeRabbit.

The effectiveness of AI code review tools is critical; Bugbot's advancement rivals the progress seen in other areas of self-improving AI for code, such as Andrej Karpathy's explorations.

Bugbot's success in identifying and resolving bugs is a key indicator of its value, a trend also observed in how tools like Bugbot Autofix aim to accelerate PRs and similar initiatives like PlanetScale's use of AI in code review.

Over 110,000 repositories have adopted these learned rules in beta, generating more than 44,000 custom rules.

How Bugbot Learns

Bugbot leverages signals from merged PRs to refine its understanding. Key inputs include developer reactions, such as downvotes on incorrect findings, and replies explaining feedback improvements.

It also incorporates comments from human reviewers, highlighting issues the AI missed.

These signals are processed into candidate rules, which are continuously evaluated.

Candidate rules can be promoted to active status if they prove consistently useful, directly influencing future reviews.

Conversely, rules generating negative feedback can be automatically disabled or manually edited by users.

This continuous learning loop is fundamental to Bugbot's evolution.

Cursor aims for Bugbot to identify every bug, requiring a deep grasp of project specifics.

Learned rules are a major stride toward this goal of continuous self-improvement.

Users can manage Bugbot's learning features and initiate backfills from the Cursor Dashboard.