2025年7月，当我们把Bugbot从测试版正式推出时，它识别出的bug中有52%在相关PR合并前得到了解决，这意味着其余的都是误报。
如今，解决率已接近80%，比排名第二的AI代码审查产品高出15个百分点。
| AI代码审查产品 | 解决率 | 分析的PR数 |
|---|---|---|
| Cursor Bugbot | 78.13% | 50,310 |
| Greptile | 63.49% | 11,419 |
| CodeRabbit | 48.96% | 33,487 |
| GitHub Copilot | 46.69% | 24,336 |
| Codex | 45.07% | 19,384 |
| Gemini Code Assist | 30.93% | 21,031 |
我们仅分析了公开仓库。对于AI代码审查产品生成的每条评论，我们使用LLM评判器来检查该评论在PR合并前是否得到了处理。
到目前为止，所有改进完全由离线实验驱动：我们对Bugbot进行调整，测试改动是否提升了解决率，如果有效就发布。
但纯粹的离线方法留下了大量训练潜力未被开发。Bugbot每天审查数十万个PR，每次审查都是一次自然实验，Bugbot可以根据开发者是否对其报告采取了行动来进行自我改进。
为了利用这些实时信号，我们现在让Bugbot能够从过去的运行中学习，将实时代码审查过程中的反馈转化为学习规则。这些规则作为额外的指令，使Bugbot的运行更具定制性，帮助Bugbot专注于特定问题、业务上下文等。
自学习规则测试版发布以来，已有超过11万个仓库启用了学习功能，生成了超过4.4万条学习规则。
学习规则的工作原理
每个合并的PR都包含一系列信号，Bugbot可以利用这些信号进行自我改进并将其编码为规则。三个重要的信号是：
- 对Bugbot评论的反应，其中点踩告诉Bugbot该发现没有用处。
- 对Bugbot评论的回复，开发者在其中解释了什么有问题，或者建议本可以如何改进。
- 人工审查者的评论，这些评论标记了Bugbot遗漏的问题。
Bugbot 将这些信号处理成候选规则，并持续针对新传入的 PR 进行评估。随着信号的积累，Bugbot 可以将候选规则提升为活跃状态，从而开始影响未来的审查。同样，如果某条活跃规则开始持续产生负面信号，Bugbot 可以将其禁用。你也可以直接在 UI 中编辑或删除规则。
我们希望 Bugbot 能够捕捉到每一个真正的错误，这需要深入理解你的代码库、开发模式以及团队的优先级。学习规则正是朝着这个方向迈出的一大步，也是我们让 Bugbot 持续自我改进整体努力的一部分。
在 Cursor Dashboard 中管理 Bugbot 的学习功能，你可以在其中启用学习规则并对最近的 PR 执行回填操作，或者在我们的文档中了解更多信息。
使用 Bugbot 自动修复功能闭环代码审查流程
Jon Kaplan
Bugbot 现在速度提升超过 3 倍，成本降低 22%，并且能多发现 10% 的错误
Jason, Yuri & Michael
在画布中与智能体创建的图表进行交互
Alex Vandak Maloney
When we launched Bugbot out of beta in July 2025, 52% of the bugs it identified were resolved by the time the relevant PR was merged, indicating the rest were false positives.
Today, the resolution rate is nearing 80%, 15 percentage points higher than the next-closest AI code review product.
| AI code review product | Resolution rate | PRs analyzed |
|---|---|---|
| Cursor Bugbot | 78.13% | 50,310 |
| Greptile | 63.49% | 11,419 |
| CodeRabbit | 48.96% | 33,487 |
| GitHub Copilot | 46.69% | 24,336 |
| Codex | 45.07% | 19,384 |
| Gemini Code Assist | 30.93% | 21,031 |
We analyzed public repositories only. For each comment produced by an AI code review product, we checked to see if it was addressed by the time it merged using an LLM judge.
Up until now, improvements have been propelled exclusively by offline experiments: We tweak Bugbot, test to see if the change improves the resolution rate, and we ship it if it does.
But a strictly offline approach leaves a lot of training potential untapped. Bugbot reviews hundreds of thousands of PRs per day, and each review is a natural experiment that Bugbot can use to self-improve based on whether the developer acted or not on its report.
To harness those real-time signals, we've now enabled Bugbot to learn from past runs, transforming feedback from the live code review process into learned rules. Rules act as additional instructions that enable greater customization of Bugbot runs, helping Bugbot focus on specific issues, business context, and more.
Since launching learned rules in beta, more than 110,000 repos have enabled learning, generating more than 44,000 learned rules.
How learned rules work
Every merged PR contains a range of signals that Bugbot can use to self-improve and codify into rules. Three important signals are:
- Reactions to Bugbot comments, where a downvote tells Bugbot the finding wasn't useful.
- Replies to Bugbot comments, in which developers explain what was wrong or how the suggestion could have been better.
- Comments from human reviewers, which flag issues that Bugbot missed.
Bugbot processes these signals into candidate rules that it continues to evaluate against incoming PRs. As signal accumulates, Bugbot can promote a candidate rule to active status where it begins influencing future reviews. Similarly, if an active rule starts generating consistent negative signal, Bugbot can disable it. You can also edit or delete rules directly in the UI.
We want Bugbot to catch every real bug, which requires a deep understanding of your codebase, patterns, and team's priorities. Learned rules are a big step in that direction, and they're part of our overall effort to make Bugbot continually self-improving.
Manage Bugbot learning in the Cursor Dashboard, where you can enable learned rules and run a backfill across recent PRs, or learn more in our docs.
Closing the code review loop with Bugbot Autofix
Jon Kaplan
Bugbot is now over 3x faster, 22% cheaper, and finds 10% more bugs
Jason, Yuri & Michael
Interact with agent-created visualizations in canvases
Alex Vandak Maloney