I feel this is a bug: whenever I’m working on a new complex plan, I tend to go back and forth asking questions and refining the plan. But for some reason Cursor oftentimes suddenly decides that’s time to do stuff: it switches to Agent mode, without warning, and begins implementing whatever prompt you last threw at it. Which is very annoying. How can I force the IDE to never ever switch to Agent mode unless I manually switch it myself?
Steps to Reproduce
Make a complex plan and keep asking questions
Expected Behavior
Never switch to agent mode without user confirmation
Hey, this is a known issue. Plan mode can sometimes read your follow-up messages as a signal to start implementing and auto-switch into Agent mode. The team is aware of these reports.
The root cause is that Plan mode has all tools enabled, unlike Ask mode. So the line between “planning” and “doing” is based on how the model interprets your messages, not strict logic. When you go from clarifying questions back to planning, the model can mistakenly take that as a cue to start building.
A few workarounds:
Use Ask mode for Q and A and clarifications. Ask mode is read-only and only uses search tools, so it can’t make edits. Once you’re happy with the direction, switch to Plan mode to write the actual plan, then start the build.
Avoid “Auto” for model selection while planning. “Auto” adds another layer of variability. Try picking a specific model so the behavior is more predictable.
Work in a separate git branch when using Plan mode, so if an auto-switch happens you can easily roll back.
There’s no setting right now that strictly blocks Plan mode from switching into Agent mode. I’ve passed this along to the team.
Thanks for these suggestions. I’m usually already staging my local changes before switching to plan mode, just to keep an eye out for any unexpected changes. Good to know that you’re aware of the reports. Do you happen to know if there are plans to fix it?