Very often recently I’ll tell Agent to do something and it merely acts like ASK mode. If I explicitly choose a model such as Gemini then its more likely that Agent will actually do the work, but if I leave Agent set as “auto” I get a response like this. Just a suggestion on the type of thing I might do.
It’s not really worth the money I am paying !
2 Likes
the problem is with Agent Auto mode. If I explicitly choose a model such as Claude 3.7 then Agent works as it previously did. However if I leave Agent set as “auto” it behaves as if it can’t enact changes, merely suggest potential solutions ( like Ask mode)
Hi Steve! This should definitely not be the case with the Auto model, as we expect it should still be agentic. Would you be able to share a request ID, so we can verify what’s going on? Thank you!
hi, here’s a request ID
5ba31c93-6dd8-45b7-a3c2-813ac8ef7059
I assume thats for one whole conversation, so there may be a moment within that where I momentarily choose “chat” mode specifically regarding where it would be normal to serve media files from ( a question)
You should spot the request above (Agent Auto ) in the logs and you should be able to see which model the Agent chose for me on that, I cant see that from here.
I can probably find some more examples in my history, or create a new one in a nice empty interaction with no context
If you’re able to create an empty one where you reproduce this issue it would be super helpful! Also, if you’re able to disable privacy mode, submit the request, and feel free to turn it off after, so I can look at the contents of the request, it would also be very helpful.
I have been experiencing this as well with “auto”, I just thought it was some kind of feature change. It’s pretty annoying and takes 3 times more requests to get things done it seems like.
Same. I’m burning through tokens at a rapid rate because each request now takes 2-3 times more prompting. Edit: looking at my prompts, I find that I am exaggerating. Subjectively, it does seem “less agentic” than it did a week ago.
1 Like
Is this a very long running task window? How frequently do you create a new chat window with a context summary?
Agents will generally start to do this in Cursor if your chat windows are too long, even with Gemini having a much larger context window. Gently remind the agent that it is in an autonomous mode and has all the capabilities/tools to conduct those steps.
Just happened in a brand new window: 890d7c2a-01b1-4bd0-8e14-a4946f2da118
I’m using “auto” mode.
Here’s how to do this:
Plan:
Keep the TanStack useMutation.
etc…
Would you like to see a code sketch of how this would look, or should I go ahead and implement the change directly?
There’s a chance it’s related to my system prompt, but it shouldn’t (for obvious reasons):>
These are keywords I may use in my prompt. If i don’t use them, ignore the below:
#talk: talk only, don’t code anything yet
#plan: make a plan before starting
#info: explain what a file or function does
Perhaps treatment of the prompt has changed?
1 Like
If you’re running in agent mode, those are likely adding confusion. If you’re wanting to use talk/plan/info you should probably designate those and stick to the Ask mode which is specifically created for those - or create a custom mode for each.
Hard to get better context without seeing some back and forth.
For example, I have a specific Architecture mode and a ruleset that the agent requests whenever it is first set to Architecture mode. Prior to that, if I left it as part of the Agentic set up and it was in my system prompts, the agent would sporadically pull from that modes rules or mix them in, sometimes it would entirely just switch into Architecture mode without me even asking. I also have a multi-chain reasoning process in my Debug Mode that I switch over to for unit testing as an example scenario.
On top of that, those are very simple and could be removed from your system prompt entirely. If you’re making a plan before you start, NLP covers you opening a task and telling the agent to make a plan - it can also help by starting in ask mode to create the plan then develop a prompt from it - and then moving it over to agent mode.
#Info is simple enough to just ask the agent what a file or function does, there is also the codebase search tool which is semantic so you can request the agent to ask what a specific file is or does within your codebase with that tool and it can infer pretty easily.
#Talk is quite literally what Ask mode is… so..
1 Like
Oh, also, by auto mode, do you mean it automatically selects which model? If so, try sticking with Gemini Pro 2.5 in thinking mode for now just to see if that helps any. I heavily avoid auto mode because its been catastrophically incompetent in my eyes.
1 Like
Hello,
I haven’t used it for about a month, but I’ve been using it frequently for the past few days. The Composer is gone. Everything continues through chat now. I checked the settings panel multiple times. There used to be a YOLO mode, and the user experience was fantastic.
Currently, I’m using it in Agent mode with Auto enabled. It does what I say and seems more stable than before, but it constantly tries to get my confirmation. I tell it to continue without asking me for approval again, but it still asks. It ends its response with “please wait,” but nothing happens. I type “continue,” it writes two lines and says “please wait” again. There’s no real action. It’s like it’s stuck in ASK mode, continuously asking me things, and this causes my tokens to deplete very quickly. I used almost 300 tokens in just a week.
Honestly, it doesn’t feel ethical. It seems like the company has set up a capitalistic exploitation system by asking the user everything and rapidly consuming tokens. I want the old setup back. If it continues this way, I’m considering switching to alternative applications.
Wow, I switched from “auto” model selector to Gemini 2.5 pro and it’s like I have my old cursor back 
I did not think that would make that much of a positive difference.
1 Like
@jdubb75 Can you recommend a resource for setting up custom modes? I’d like to download your brain and understand why and how to use different modes.
Are you in auto-run mode as in the replacement mode under features, or are you saying auto mode as in it automatically selects your models based on the task?
If in auto-run mode, add a rule to inform it that it’s in an autonomous agentic mode so it gets a little reminder gently at every response. Ensure you’re not allowing task windows to run for too long to maintain better coherence and quality as well.
Finally, if you are in auto selection mode for models, turn that junk off - its terrible and always has been. Just stick with gemini 2.5 pro 03 - NOT 0506 and you should see better results.
Happy hunting
Best I can do is recommend to break logic down and systemize how you approach your projects! Sharing mine may not be the best fit for all your situations and I program in specific languages/only tackle certain types of projects.
My methods may end up being overtly mechanical and rigid for some users to deal with. For reference, I have an ‘Audit’, ‘Architecture’, ‘Debug’, and several other modes that I’ve set up but they all follow different processes and leverage certain MCP tools or guidelines that I’ve created specific to my domain
Very glad it feels like you got your old cursor back!!! I know that feeling very well and I’ve became used to having to navigate all of the spontaneous updates to maintain that familiarity.