Hey, thanks for the report. I understand your concern. The agent really should stop once the task is done.
This is a known issue, and the team is working on improving how the agent detects when a task is finished. Similar cases were discussed here and here.
Temporary workarounds:
-
Use the Stop button actively. As soon as you see the task is done, click Stop. This should prevent extra costs.
-
Set up User Rules (Cursor Settings > Rules). Add a rule that clearly limits the number of iterations or blocks extra verification. For example:
- Stop immediately after the task is done
- Don't rerun browser tests unless I explicitly ask
- Max 3 iterations per task
- Use more specific prompts. Clearly say you want one quick change without extensive testing.
Could you please share:
- Request IDs for the problematic sessions (Chat context menu > Copy Request ID)
- An example prompt that triggers this behavior
- Whether you have any User Rules set up
This will help the team fix it faster.