The most annoying issue is that Agent is by default already “too active”, implementing stuff , doing code changes , nobody has asked for.
Especially annoying if you ask what option there are to appraoch a certain problem, then it presents the options, and instead of letting the developer read and digest them … it just moves ahead "I think Option x is best , so I implement …
Yes, you can tame it a bit by giving strict rules, which you might have to repeat as it sometimes forgets/ignores them, but the default should be “wait for instructions how to move on”.
I honestly couldn’t agree more that all models’ performances have degraded significantly, to this point I’m questioning if I should keep my subscription or start looking for alternatives.
I’m counting on cursor to help with my daily work having it fail isn’t an option, unfortunately, if 20$ isn’t enough please offer higher subscriptions but don’t degrade the model without a disclosure
Yeah, honestly it got to the point where I have a feature request. I have paid additional for the claude 3.7 sonnet max where i set a spending cap and it charges me per request. But honestly I feel like an idiot when its going in circles hallucinating and I have to keep arguing with it because of stupid mistakes and not properly paying attention to the instructions. I feel like I just had money taken off my pocket for nothing. Maybe you can add a feature to seamlessly request a refund for the past “n” requests where the model just kept hallucinating. This is ridiculous. At least for the “max” model this should be available.
Its a bit ridiculous to think that Cursor would degrade the model performance on purpose. I mean why would they do this if they would lose customers?
Instead why not look how you can improve your prompts?
lots of people post here without any details what they are trying to achieve or asking others for help.
Others expect miracles or to be refunded 20 USD if they continue to submit prompts that fail until they run out of tokens.
We all have sometimes issues and make mistakes, and AI sure does make mistakes.
I wrote the other day rules for my project where it suddenly broke code output. Not because of cursor but because the whole setup has to work with many details and sometimes too many details are bad and sometimes too few details are bad too
this 0.47.8 update is so full of bugs, it’s barely usable, cursor team please look into ti. I cannot work at all! If this is the way forward then would likely switch as it’s really competitive right now.
I am literally paying only for the tab tab tab auto complete. It is so much faster and better than the one from Github Copilot (including the new next edit suggestion feature). I wish I could pay for just that feature. I don’t use composer or agent. It would mess up too much. It will also never be better than using roocline etc directly. The only stand out feature for me is the tab tab tab and that has increased my productivity and speed by a lot. Love it!
Same issue as the others here, using 0.45.14, and the agent mode has been performing worse and worse. Considering stopping the subscription as i spend as much time fixing code than it would take to write it.
Examples of bad performance for the agent:
not taking into account existing functions to re use them (even when specifically told)
not taking into account existing functions to have the same style/standard/framework (even when specifically told)
not functionning code when making simple functions
This is all very interesting. I am using it for various project, Dart, Python, PHP etc, I have been having great success with it. However, I do treat it like a junior developer with very specific guidelines.
Maybe the Cursor team can bring out documentation for setting yourself up for success. It is a great pitty that people are having so much negative results. Just when we thought things is taking off with AI assistance development.
After the update, this 3.7max. It’s like a mentally retarded person. The more it is updated, the worse it is. I don’t understand code at all. When it was version 3.5, I could use these tools to develop various telegram robots. Although it was very laborious, I completed them all.
But this latest version. The various instructions given to him were not followed at all. Because they were not followed, I asked him, if you don’t listen to me, what should I do? You tell me the instructions. Then I sent him the instructions he sent me in reverse, and he also said that he understood, but he still didn’t do it, and still didn’t check the complete document to fix it. Then it kept going in this dead loop. In just two days, $300 was gone, and nothing was done.
I have just downloaded with the intent to purchase pro. Until I saw this. It is very concerning to see this within 5 mins of logging in:joy:
Should I go pro or find an alternative?
Please. Any assistance would be appreciated. I can only afford 1 shot at this unless i want to start sacrificing meals.
Please help. Please advise.
Have you exhausted the trial? Maybe use the trial if you still can, maybe things will level out by then??? Maybe… My experience hasnt been good over the last month
You should use the trial first to get some experience. It helps if you have experience in prompting and writing rules. While I use AI for 2 years now it does require me to prompt differently when using Cursor.
running this command on my iterm terminal for the project I’m launching cursor --disable-extensions . and when Cursor is up, renable and the extensions and try to start a new chat right away; worked for me!
I was initially just going to like the other replies in this discussion, but I agree so strongly that I felt compelled to share my own thoughts.
I agree with this quote from Original Post: "
I have been using cursor for a long time and when I used version 0.44 and 0.45 I didn’t have any problems, on the contrary, I was happy to work with it. But from 0.46 to 0.47 I have been experiencing a lot of problems with cursor every day…
"
We all understand that software development is challenging, and our intention isn’t to criticize the developers personally. However, as users, it’s our responsibility to provide honest feedback about our experience using Cursor.
It concerns me when a pinned topic titled “Community Program — We Want Your Input!” feels more like a prompt to list only positive aspects. Meanwhile, Cursor’s performance has been inconsistent across versions, and the recent decline has led to excessive wasted premium credits.
I feel like I have to treat Cursor as if it’s deaf-blind—carefully pasting full code snippets—yet it still claims the code is cut off. As a result, it generates a lot of unnecessary lines, seemingly ‘taking shots in the dark.’
Performance has noticeably declined over the past few days, making it exhausting to implement suggested fixes that actually work as intended.
Cursor’s agent feature flat-out ■■■■■. It causes more unpredictable issues and wastes way more requests than necessary. Honestly, it feels like it was designed this way just to force us to stop using the original edit tools and rely entirely on Sonnet.
