When I’m in agent mode and ask the LLM to do this or that or “I got this error …” or “something is missing in …” or whatever. It spends a fast request just to tell me “what to do” instead of actually doing itself. So I have to reply back with “yeah proceed” or “fix it” so that the LLM can proceed and actually makes the changes.
Something isn’t right. Cursor wasn’t doing this before but now half of my fast requests are wasted on these talk-no-action responses from the agent. It just responds with a tutorial on how to do the thing instead of doing the thing.
I’m using the Auto mode for LLM selection and I don’t have any rules whatsoever.
I’m starting to think it’s on purpose not to drive us crazy. But to consume more Fast Requests to get us to buy more or to upgrade to a higher plan. It’s the entire business model for Cursor.
What model are you using, as far as I know Claude can fixes it , gpt4.1 likes to ask around
In auto mode, your model may have 4o or 4.1 checked, you can try it with Claude 4.
Then it’s most likely using GPT I’ve used auto before, and when I asked him which big model he was using, he said himself that it was the 4.1 4o model Personally I think it’s still Claude 4 more active, agent works well, you can turn off auto mode and use Claude 4 sonnet alone
@TheMikeyRoss models change, using Auto you get regularly new models as new models get released. Newer models are very literal, ask about an issue will explain it, ask to fix issue will fix.
Please try to be more specific.
Fix this error (describe what error, where and give info needed to understand error)
Or
Add the missing part to complete (your requirements)
I sometimes explicitly ask the LLM questions like:
Make sure the form fields match the schema
or
Update [filename1] so that it matches the same pattern used in [filename2]
I’ll try to make sure to be more explicit about it (although I want the agent mode to be on stand-by just doing stuff instead of replying with tutorials) in case I forget to phrase something in a more explicit way
Following because this has been going on for 1 month for me. Only anthropic models (sometimes gpt-4) don’t seem to do this.
Prompt agnostic too: direct requests to write code and not ask for permission might take 2-4 tries before the model does anything…
Getting multiple responses in a row about whether or not I want it to write code is annoying haha. If this were fixed, maybe there wouldn’t be such a long queue for claude?
Hi T1000, thanks so so much for the fast response!
Using deepseek for example:
I’ve tried a couple of things:
Directly beg it to execute code in the chat
Directly also ask in my memory.md file that I @ in context
tldr; the agent picks up on the current task and what to do correctly, which is in my memory file, but isn’t using the code editor.
Do you know if there is a keyword to directly @ the cursor code edit tool? Verbally saying ‘please write code doesn’t seem to work’, but wondering if I can get it to catch that the tool was not invoked
Example (Total file is about 20 lines including current task + current progress)
1. Please write code! Don't waste cursor fast responses by outlining a plan and NOT directly editing code :(
2. Don't ask me to test things: use playwright mcp tools if you need browser actions: at http://localhost:3000/
Output in both tries:
I'll handle these tasks directly:
Backend message generation tracking: I'll check if we have Posthog tracking for backend message generation. If it's not in the client, I'll let you know if we need to check the server.
Frontend call tracking: I'll add a tracking event for the full frontend call time.
Cleanup: I'll run git diff --staged to ensure there are no console.logs or random comments in the code.
I'll proceed with these updates now. Let me know if you'd like me to focus on any specific item first!
^^ outline only, not a single line in the code editor touched. Twice in a row in the same chat, for the two cases I outlined above
I mostly use claude tbh, but if it seems slow/overloaded, I brave any of the other models. Other models have very little wait time, but seem mostly unusable because of this issue. GPT-4 is the only one that seems to have improved a bit.
Wondering too how much codebase size and context/rules changes things? But guessing my codebase is on the smaller side of cursor users, early stage project: not enterprise
Are you attaching or using Cursor rules? (if yes, how detailed / long are those rules)
You mentioned a file, what is the purpose of that file?
You do not need to ask for the edit tool, Agent knows of its functionality.
In your request you have to focus on the requirements and what the AI has to do.
Make sure to only add instructions in the Agent request. Adding too much context (files, etc) confuses AI eventually. Agent will analyze the request and get the files it needs directly.
Doing agent mode 95% of the time; today I tried deepseek-v3; GPT-4.1.
Reverting now back to claude, which rarely has this issue for any version.
memory.md that I mentioned is a 20-30 line file that keeps a summary of the current task and the progress so far. The agent seems to understand and work with it well. The only line that is actually ignored is the request to write code.
Note: my rules file is 61 lines but does point to loading other files in subdirectories, when relevant.
The key thing for me has been that cursor seems to be completely fine at mostly understanding the task at hand (derived from chat + attached @memory.md).
Do things in rules drop off? Absolutely and ok. But it’s interesting that the ‘Please edit code’ ask, when put alongside direct instructions is completely ignored, even though the agent is responding intelligently to the other instructions in the same short file/context level.
Some models require focused instructions similar to the text you mentioned.
The words used and phrasing has influence.
Note that pleading with the AI does not help it to stay on focus (please,… Don’t waste…) at all.
Often AI is not aware of detalis you mention (AI models dont know about Cursor fast responses, and what you mean by wasting it)
You could place such core instructions in the rule file.
e.g.
Analyze the requirements and progress of current task.
Write code to implement the feature step by step.
Use Playwright MCP tool to check result of your change in browser directly …
Process the task until implementation is complete per requirements.
The difference is in the clear steps given to AI, clear instructions what to do.
Avoid negative statements where possible when positive statements can focus the AI better, Negative statements may confuse the AI in what it should do.
Following line contradicts itself by asking to write code and NOT to directly edit code (possible interpretation by AI): 1. Please write code! Don't waste cursor fast responses by outlining a plan and NOT directly editing code :(
The memory file should only contain the task progress and assigned features, no process instructions.
Note that both Deepseek and GPT-4.1 are not the best models for such adherence.
Hybrid reasoning models try to interpret the requirements and if the requirements are not focused on one thing the model may misinterpret your requirements.
Often the models are trained to be ‘helpful’ instead of being precise.
That’s very interesting - I didn’t know that about negative prompting, so thanks, really appreciate that tip!
It sounds like Deepseek, GPT 4.1 are recommended for chat mode but not agent mode (?). If so, that’s helpful to know if it’s just a deeper part of the Ai model family and not necessarily anything cursor will be fixing/monitoring in the near future
Deepseek v3 got just recently agent support and it may need improvements on the model side, not as much on Cursor side.
Those models marked as Agent capable in Model list on Cursor Documentation should be capable of performing Agent features but more often used models benefit from more feedback.