Since the 3.11 update, I’ve been using BugBot locally in Cursor, and I’ve become a huge fan of this feature. But one thing I’ve noticed is that BugBot consistently produces much higher‑quality results compared to Agent Review — clearer reasoning, more accurate suggestions, and generally more actionable output. And Agent Review often produces nothing at all (literally no output).
For example, you usually will have this with BugBot, very clear.
What’s interesting is that BugBot only consumes Composer 2.5 quota, while Agent Review uses my API pool quota. Despite this, BugBot feels significantly stronger in practice.
So I would like to know why. Specifically:
Are BugBot and Agent Review powered by different internal pipelines?
Do they use different prompting strategies or agent architectures?
Is BugBot intentionally optimized for debugging while Agent Review serves a different purpose?
I’d love to understand the actual design differences between these two tools, and why BugBot’s output quality feels noticeably better. Is there anything to do with agent eval? Is this because BugBot is a special subagent per se?
Any insight into how these two systems differ under the hood would be greatly appreciated.
Hey, thanks for the detailed breakdown. I can see both screenshots, the chat flow Run Bugbot Review with rich output, and the in-editor Agent Review with no issues found. You called it correctly, and this is expected behavior. These are two different pipelines, not one feature in two places.
Quick answers to your questions:
Different pipelines? Yes. Run Bugbot Review in chat runs a dedicated Bugbot sub-agent inside a normal agent session. It explores the codebase more broadly (your screenshot shows Explored 27 files, 16 searches) and renders findings as detailed markdown. In-editor Agent Review is a separate, lighter pipeline built to run fast and automatically on every commit, so it’s scoped to your diff and intentionally conservative, which is why you often see no issues found.
Different models and quota? Yes. The chat Bugbot sub-agent runs on Composer 2.5 Fast, so it only uses Composer 2.5 quota, like you noticed. In-editor Agent Review uses a different model and is billed by tokens from your API pool. That’s by design, not a misconfiguration.
Different tuning? Yes. Agent Review is tuned for speed and low noise so you can run it automatically without spam, so the output is more restrained and short. Chat Bugbot is more exploratory and naturally more verbose.
If you want a deeper pass directly in the Agent Review panel, check the Deep Review option in the dropdown. It’s closer in thoroughness to the chat flow.
