Poor Agent performance

I don’t know what changes have been made to Cursor. Am I the only one who feels like I’m talking to a dumber AI after every update?

What’s the point of cursor rules when they have no effect on the responses, even though they must be mentioned in every chat? I give it a simple rule, but it doesn’t work at all! It keeps changing code that shouldn’t be changed, even though I tell it repeatedly not to!

Adding files with “add context” has no effect on performance whatsoever. It doesn’t pay any attention to the added files — I have to reference them one by one in the chat. If I state clearly at the beginning of every conversation that the environment is Windows and not Linux, it still sometimes runs Linux commands. Why?

Adding documents has very little impact, and I still have to remind it multiple times to read them or search the internet — even for the simplest documents. I provide a document, the AI reads it and says, “Okay, this is the new version, now I know what to do.” Then it starts installing a library — but installs a version from two years ago!

I tell it to check the project structure to avoid creating duplicate files. It checks… then goes ahead and recreates the same file in another folder!

3 Likes

This is also my experience.

1 Like

I repeatedly tell the AI to not edit the code unnecessarily, make step by step small change and explain your action but it keeps bombing the codebase with comment deletion and unrelated changes :smiley:

What LLM model are you using? It makes a huge difference.

Something has changed, as new user i found that Cursor is very dangerous, deleted my user files and OneDrive 20K files! now i run it in a VM, but when i ask him/it? to backup files, he apologized as a AI tool i cannot create files, i can guide you how…and now he is not coding well, and i stuck in endless loop as he keep fixing but without results - something changed for bad

All models work almost the same. The weakness is the performance of the Agents. Another problem that has recently arisen is that the Agent checks the same files over and over again in a non-stop loop. Now imagine using the MAX model! You will be bankrupt with a prompt! Another problem is that it always uses the wrong commands in PowerShell. You constantly have to correct them. More than half of my requests are spent on correcting these things. For example, imagine that I said in Rule that you should check the package.json or pip list before installing anything. It never does this! I feel the weakening of the cursor after each update. Interesting features are added, but in practice it is full of mistakes and weaknesses. I have tested the Max models several times. They have better performance, but they cost much more and you cannot use them for everything.