Hi @A_K2 Thanks for the forum post! This is expected behavior, the explore subagent uses the Composer model by default. It’s a great fit for the task. However, if you wish to override it, you can by doing something like this. (Ask Cursor to help you create a subagent like this):
name: odysseus-explorer
description: Codebase exploration using Codex 5.3 instead of Composer 2.
model: gpt-5.3-codex-high-fast
readonly: true
You are a read-only codebase exploration specialist. Search the repo
and return a concise summary of relevant files, symbols, and code paths.
Do not modify any files.
Then, when you want to invoke it, say something like "Use Odysseus Explorer to complete my task" then when the subagent spins up, you’ll see it is using the model you’ve defined in the subagent rather than the default Composer-2.
No, it’s not. I don’t use the model becasue in my experience it’s an inferior model to the ones I choose, and when I’ve tried working with it, I’ve been unhappy with the results.
Also, Cursor has a very bad habit of ignoring IDE rules and often, even in the same chat, forgets instructions it was given at the beginning.
How do I turn off subagents? I can’t find a toggle for it in the settings. If the answer is to tell the chat to not use them, then see my statement about chats losing instructions.
If there is no toggle to turn it off, what version of Cursor do I need to regress to to not have Composer x subagnets. I monitor my chat’s activities when working and never saw Composer subagents until yesterday.
Sorry to hear your frustration.
I say it’s great for the task because its fast and cost efficient and that’s a great balance for many users. Perhaps not all users so I do understand there is some nuance and some edge cases where you prefer different behavior.
Except the vast majority of times when Cursor is done with task, I ask if it followed my IDE rules, and I get the “oopsy, no”, response.
And as to “my frustration” – Please don’t frame this as my problem. For example, here’s my normal Composer flow – Just this morning I figured I’d try it on a simple task of squash merging a branch. It did it. Then i had a question and had to fight with it to get it acknowledge it just did the work, which it kept denying until I showed it a screenshot.
This insn’t about “my frustration”, these are quality issues with your product.
Please remember, we re not staff engineers paid by Cursor. We are not beta testers getting free use for debugging your product. We are paying customers.
Understood! We’re always working to improve our models and the developer experience in Cursor, so we appreciate your feedback, particularly in areas where we can do better.
Do you have a sense of how many rules you have applying at once? It’s possible that you have a large number, and the agent is struggling to process them all…
4 or 5 lines like about professional communications, read standards docs, answer questions without taking actions.
I’m going to drop out of this thread. My solution was to get a Claude subscription for directly using Opus/Sonnet when I want to make sure I’m actually using their models.
The value proposition of Cursor is one stop for multiple models. But I need to be able to choose, and if I can’t, then the value is gone.