If I am having a terrible time with Cursor, 9/10 times it’s because the agent switched from my current model selection, i.e. gemini-2.5-pro-03-25 to Auto.
Auto, to be completely frank, is garbage at the level of intensity I push in a given thread. I lose at least tens of minutes and ge irritated because Cursor is changing my default agent decision.
I curate a big, perfect thread with Gemini over the course of days.
I go to sleep
I come back to pick up where I left off, and the model has switched to “Auto”.
I don’t notice, so I use Auto for maybe an hour. It pollutes my thread with GARBAGE code, false assumptions, and arguments.
I notice that it switched to Auto at some point. I can’t go back at this point because I’ve already changed too many lines of code.
The REST of the thread is GARBAGE because the thread is polluted.
My day is filled with high cortisol levels.
This is objectively a bad decision. If I have to go through this again I’m switching to Cline and that’s going to ■■■■ hard. But I don’t need cursor. I need an agent that uses Gemini 2.5 Pro. And I’ll pay for it. I’m paying for you to call that api for me. Don’t get greedy.
Your problem does not happen to me ever, on 0.45 all the way up to 0.50. If it does not work I suspect you might have to reinstall cursor, as there might be a problem with cache on your local machine (happened to me with similar problem).
The problem I see with the “Last used mode” default is that it doesn’t really speak to the issue. The issue is that the default model within the Agent mode is defaulting to Auto, instead of gemini. Setting the default mode to “Last mode used” will just set the mode to Agent, which is the current behavior.
I have reinstalled the most recent version of cursor.
Right now I’m testing out using a custom “mode” even though the custom mode model selector is broken. I have yet to see if the custom mode reset to “Auto” model and it might take a number of days to confirm.
I appreciate your help. I’m gonna see if my new mode works for a couple of days and if it doesn’t work then I will try switching to the last used mode. If that doesn’t work then I will reinstall cursor again.
+1. Every time I wonder why the answer is so bad it takes me a minute to realize it autorouted the request to some random model. For me the temporary workaround is to disable all model access but 3.7.
I understand you need to balance performance with cost but this is so clearly not it. If you’d like to compromise do it like you did it with rules - some prompting is done under the hood, but my explicit prompt is the ultimate oracle. Same here, perhaps give me suggestions, but ultimately let me set the right model for the job and don’t override it without my permission.