Hey Cursor Team,
I have been happily using the new agentic feature of 0.44 together with Claude 3.5 Sonnet since Christmas.
However, today I noticed a significant decrease of its performance, it became completely dumb, lazy and was not helpful at all. I am not sure what you have changed behind the scenes but this should be reverted.
Hey, have you tried starting a new session? It usually helps when the context window overflows and the model starts hallucinating.
Yes I start new session quite often for that reason. It was a really strange day with Cursor today, it even forgot to implement simple Python functions and just wrote … as the implementation.
Hope it goes back to being smart next week with 0.45 
I feel the same pain. I downgraded Cursor multiple times from 0.44.11 to 0.44.3, thinking that it was the problem, but the problem is definitely Claude. I’m using sonnet-2024-10-22; I haven’t tested the older version yet.
I’ve found that occasionally claude can get into an unproductive mood. I had one instance, that I wish i screenshotted, where claude said it was going to do something and then ended the agent call. It did this for about 6 consecutive prompts.
claude: Ok I’ll make that change -ends without making a tool call-
me: you didnt actually do anything
claude: you’re right, let me make that change now… -ends without making a tool call-
etc
I ended up asked it if it was doing this intentionally, i think i used the phrase weaponized incompetence and it told me “to be honest im just an LLM and i dont have the capability to write a program like that” and then i gave it a pep talk and it worked fine after that. I’ve also found claude does a lot better when you say “we need this” instead of “do this” or “i need this” and it continues having better results when you thank it and congratuate it when it does something right. Kinda weird but I get the same results with humans and its trained on human data (i think) so…
Granted they do seem to be having high load issues between the cursor and anthropic servers lately. Its really really bad to be on the slow queue now. You end up waiting 6 minutes for it to return something useless. Agent is pretty much unusable if you’re on slow pool right now. I wouldnt be surprised if they were trying to tweak it down to use less tokens
Good to hear that I am not alone! Probably they are optimizing for less token usage behind the scene to get around the limits. I really hope this stabilizes, so we can rely on the code quality more!
Maybe I just got a bad roll, but today Cursor Claude seems particularly lazy.
Instead of fixing my compile issues correctly (Swift) … It instead decided to simply remove the dependencies and “improve” the logic by “removing unnecessary complexity” thereby completely changing the entire purpose of the function, JUST to solve the compile error.
And yes, I always have in both Rules for AI and .cursorrules “always make focused fixes with minimal code changes, never remove or alter existing functionality…”
After calling it out on it, it said “You’re absolutely right. I shouldn’t have modified the core functionality and instead focus on fixing the actual issue. Let me restore the important functionality…”
So just ended up wasting a call… back in the queue line.
It wouldn’t be half as frustrating if the slow pool hadn’t become so insufferable as it has these last couple days.
Hey, my best recommendation here is to regularly start a new Composer session, as this helps to keep the LLM focused and on topic with what you are trying to do.
Long composer sessions sound like a good idea as they are full of context and conversation about your codebase, but Claude can get overwhelmed and confused with a lot of it, and almost ends up not bothering to do the right things, even though it hypothetically had all the tools for the jobs (context, knowledge from your chat, etc).
Even if you say I need to get this coded, if I don’t my boss will start killing my puppies, by putting gas on them and lighting them on fire. I only get a slightly better result in anthropic api console when I do that.