Hi team,
I’m having issues with Cursor responsiveness, agent quality, and wasted prompts.
I’ve been using Cursor Pro for a while now, but its performance in recent weeks has become really frustrating last few weeks.
-
No response after prompting:
Cursor often gets stuck on “Generating…” and never finishes. I’ve had to cancel and retry multiple times, but even then, the prompt may fail again. This happens far too frequently.
-
Agent doesn’t apply changes:
About 80% of the time, the agent just describes what it wants to change instead of actually modifying the code. I then have to issue a second prompt just to ask it to apply what it already suggested. This is both inefficient and a waste of tokens/prompts.
-
Agent-generated code has bugs:
When I run into issues with code that the agent itself just generated, it often blames the bug on “my existing code” instead of acknowledging its own mistake. Again, this leads to another prompt just to fix what it broke.
-
Fixing one issue introduces another:
After fixing bug A, bug B appears. Then asking it to fix B sometimes reverts the code back to a buggy version from A. It creates a frustrating loop of back-and-forths, wasting time and prompts.
I hope these issues are being prioritized
Please let me know if there’s any upcoming fix or workaround.
Thanks.
13 Likes
Seeing the same issues in terms of performance. Have also started a thread on this. Potentially victims of their own success. I can’t tell you how much I love this product, but it has to remain effective. My fix for response times is to stay on Auto, but of course, with no thinking, the results are not going to be anywhere near as good.
Fingers crossed the dev team find a lot more bandwidth for us to play with.
2 Likes
Also getting a much greater amount of connection failures, which is ironic seeing as I just had 1Gb connection put in! 
1 Like
I have the same frustrations - it used to be a great tool, nowadays it is as good as prompting in Claude directly and using VsCode.
The agent not applying changes and describing what it thinks we should do is just poor compare to before.
I get things like:
The agent should know if I have those files in my project or not…otherwise is not an agent, just another AI chat.
2 Likes
It’s still subject to context saturation though I guess, so unless you create a persistent source for info like that, it’s eventually going to forget that part of the convo, as it’s not going to poll the project structure on every prompt. I’ve started using project templates that try to avert the issues of saturation, if that’s of any use to you?
1 Like
My Cursor stopped working since this morning. No signs of bot liveness, lol.
I’ve switched to using Windsurf instead. It felt like a decade ago when Google went down!"
2 Likes
At this point jump ship to the surfers and never look back again. However, all these IDEs will converge at one point offering almost identical (if not exactly the same) features. It’s a matter of time, so if you need results NOW try the alternatives, otherwise, wait for Cursor to ship major improvements soon.
I just came here to say the same exact thing regarding AGENT.
It only suggests and then requires me to tell it to continue, whether it be:
“sure, sounds good, fix it”
“please find the files in my folder structure and modify them”
“yes, please, continue”
What happened to AGENT?
2 Likes
It’s more than inefficient. All the precious details from the thinking process are gone from the context, when the actual implementation of the planned idea is delayed until next prompt.
+1.
No response at all, unsure whats going on and then when it does respond, it tells me the change to make, even though its in agent mode.
2 Likes
In the span of the last 2 days, Cursor has become entirely useless. Any agent interaction results in the program completely freezing and it requires a force quit.
Also, It totally changed on a dime regarding it’s chat code application. Now It’s just suggesting what to do instead of showing you acceptable diffs.
This is remarkable because all this came out of nowhere.
I reinstalled it, too.
It is straight-up broken.
