Gemini 2.5 Pro in Agent Mode cuts off responses too early

Bug Reports
1

In Agent Mode, the gemini-2.5-pro-exp-03-25 model tends to cut responses short when handling slightly more complex requests. Instead of continuing until the task is fully completed, it stops prematurely, possibly to avoid making too many tool calls in a single request.

2 Likes
2

Can you use the supported version?

gemini-2.5-pro-exp-03-25

Edit: corrected model name as it existed at time of writing that response.

3

gemini-2.5-pro-exp-03-25 is the default model name retrieved from the Cursor servers.

1 Like
4

Can you please read the page there? I am using cursor as described with that model.

in worst case you can always remove that model from Cursor / Settings and add the correct name.

5

image
image1098×248 8.52 KB

I have the Pro plan btw.

6

I sincerely apologize i misread you comment.

Do you use an API key in settings?

7

Nope

8

Ok thanks, in this case the Cursor team must take over as I dont have that access.

Which cursor version was it?

Amd gemini-2.5-pro-exp-03-25 works in general but not well?

hm sorry wrong screenshot that i posted again

9

image
image516×730 44.9 KB

Yeah, the gemini-2.5-pro-exp-03-25 model works pretty well but it tends to stop when handling more complex tasks like analyzing an entire codebase. Even if I specify that it should continue the analysis without asking, it still pauses at the end and prompts me to continue with the next file or folder.

10

Right, from what I see on my machine its overall giving good responses but yes sometimes the tool calls are a bit an issue.

Not sure if this is something Cursor is working on but the model is new and they did yesterday the Agent integration.

Do you have a Request ID they can look into?

11

Yeah, I just clicked the ‘Report Issue’ button and submitted a report, although I think Cursor is already working on improving it as we speak. Like you said, it’s a new model and still needs some adaptation for Agent mode.

1 Like
12

I think so. in agent area claude 3.7 sonnet would be better

13

Hey, we’re still working on integrating this model into Cursor. I think we’ll soon optimize its performance in Agent mode. Thanks for the report.

1 Like
14

Why?

15

Apologies, corrected my above comment.