I hate to say it, but 0.46 seems to perform worse than 0.45. After updating every day hoping for 0.46 to improve, I’m throwing in the towel.
The UI changes are a step in the wrong direction
Only one chat is an obstacle, I don’t like that its agent by default with no option to change the default.
Agent is remarkably worse
The new agent produces really poor results and I find that I now only use ‘Edit’ mode. In 0.45 I typically used Agent mode. It is over-eager with its tools, and spends 30 seconds of every request calling tools trying to ‘understand’ the issue, but all its doing it diluting the context I have provided with low quality and often unneeded data.
It frequently and quickly loses track of the request and starts making other changes.
It ignores course corrections, once it has made a mistake once, it keeps returning to the same mistake again and again.
All this and I haven’t experienced any actual bugs, just a overall decrease in the reliability and helpfulness of the AI features. I’ve downloaded the leading competitor IDE for the first time to see if that is any better.
I am having the same issue after updating to the latest, it just gets stuck and doesnt do any file changes. something major broke in this release, can you please tell how to rollback?
Version 0.46 is incredibly bad! I got so frustrated today because it kept doing stuff I didn’t ask for, breaking things and removing content along the way.
Switched back to 0.45 and it’s like a breath of fresh air.
Good luck preventing updates. Mine auto-updates every time the app resets even with the Update:Mode set to none. I’ll just re-install it every time, its still a better process.
I just rolled back to V0.45 and it feels so much better again lol. cmd+l for a fresh chat window everytime and it always brings in the context of what I have highlighted.
Rolled back to 0.45 as well. The code generation UX is also frustrating. Have to wait for the model to generate all the changes and apply them to code before I can start verifying the output. Previously I could see what the model is generating and stop it in between if it’s going in the wrong direction.
I am having exactly the same “diluted context” problem - the main agent seems to do all the thinking and planning with the files I tagged, then as soon as it does its subroutines, it starts reading the files again as if it was told to do the task alone - it’s terrible. And I’ve historically loved using cursor for this type of stuff
If I have a side-question I don’t want to add irrelevant context by asking it in the same thread. Switching the mode forces me to keep everything in the same context, or switch between threads by using the history. Both options require more clicks and more managing. I liked being able to have my agent run and ask chat questions in a different thread at the same time. You could ask questions in Chat while the Agent was still running in the background. Sometimes I’d work on two features at the same time this way.
I totally agree. The agent starts every request by trying to ‘understand’ the code by searching within the file I’ve already shared four times, searching other parts of code that I know are not related and then it runs off trying to use the cli to do things like check the file structure and see if files exist when I’ve asked for a change that should only happen within one file. I use Typescript or strictly typed Python. If I ask for a change within a function, it should not need any context outside of that function except maybe to lookup the types. All the context is there already! Thats why we use types in the first place!
I’ve really enjoyed the Cursor experience up until this point. My experience with the competitor product was underwhelming. I found their agent performed better in some cases, and worse in others. Cursor is still probably the top product, but won’t be for long at this rate.
• Agents are automatically modifying code without being asked. Even when it generates incorrect code, I have to fix it, but then it deletes my corrections and reverts to the incorrect changes it made before.
• Sometimes, it ignores MDC and generates random code, even though I explicitly asked it to read MDC, but it completely disregards that request.
• Requests always start by reading a large number of files (even unrelated ones) and then get stuck or ask to “continue” because they exceed 25 interactions.
How to use in any mode without agent and be able to apply code section manually? I rolled back few days ago to 0.45 but today got automatically updated…I hate it. I start to believe that cursor is just on money grab and wanna you use much more tokens and your allocated credits so that you will top up for more credits.
Agent in 0.46.10 tends to lose track of what we were doing. Keeps going in circles and draining request quota. Lost whole weekend trying to make it do the simplest thing. Terrible experience.
Cursor team should provide a way to reliably downgrade. Or in general switch to a particular version.
I have been testing 0.46 and it really is awful. I have reverted to 0.45. The agent mode keeps doing stuff I didn’t ask it too and it is generating much worse code than before.
Not a fan of the new chat interface too, the tabbed ‘composer/ask’ felt much better, as I could switch between composer(agent) and chat(ask), without losing contexts and previous answers.
