The agent just doesn’t seem to want to do the work since the upgrade to 2.0. It delegates it to me to do. I have to have additional instructions to get it to do anything. It has become overly cautious. Maybe we should have some kind of setting that allows us to how “proactive” we want it to be. I also miss the emoticons and bold titles because they gave me something visual to key off in these largely monotonous conversation logs.
Steps to Reproduce
Debugging code
Expected Behavior
Pre 2.0 behavior
Operating System
Windows 10/11
Current Cursor Version (Menu → About Cursor → Copy)
I also find rules to be inconsistent so I’m using them less also.
I basically have a library of commands that I use and I put everything in the command - it respects the commands very reliably.
The issue with the rules and memory is it “decides” what is relevant when it starts the chat… but then as you chat it doesn’t reliably recheck what’s relevant as the scope changes.
I also try to start new chats very regularly as the longer the chat the less reliable it is to follow the prompt
on memory - so interesting - I have tried, but would agree that it doesn’t work.
Agree on rules to be inconsistent - @cursor people please fix this, or make it more obvious in the chat that our rules are actually being applied.
@mitch_dev - what do you mean by “I basically have a library of commands that I use and I put everything in the command - it respects the commands very reliably.” What is this you are doing here?
Agree on : “The issue with the rules and memory is it “decides” what is relevant when it starts the chat… but then as you chat it doesn’t reliably recheck what’s relevant as the scope changes.”
On sessions : “I also try to start new chats very regularly as the longer the chat the less reliable it is to follow the prompt” - I do this all the time as much as I can and just commit even if the code is not working.
Bottom line Agent 2.0 has become an absolute pontificator, suggesting that I do everything. For me after using this auto mode this week, I think it is a step backwards and left me more frustrated than ever and now I am turning to alternate models again. The auto mode we had before 2.0 was way better. It’s weak, it’s way less confident, it’s way harder to work with to get it to do what you want it to do. I was having far greater success with pre 2.0, especially with solving complex bugs.
You can create custom commands which can either be project specific or global. Since I work on many projects I have a library of global commands. I get AI to create them. I have a command to create commands actually.
Anytime I notice something or learn something then I ask AI to add it into the command. I split the commands up into specific tasks such as:
git commit
git deploy
database change
database rls
audit codebase
audit database
implement feature
fix bug
etc
I NEVER chat with the ai without a command present - the command is the driver and I just give additional context.
All new features basically follow this process:
plan the feature with my plan command (new chat)
creates an instruction file with the finalized scope - splits it into phases so I can run multi agents at the same time
implement feature with the implement command + attach the plan file (new chat for clean context)
summarizes the scope, rechecks codebase to ensure its still valid etc
I confirm the scope again or make final adjustments
it implements
stops if it needs database changes/rls
I confirm with the database command
I test it
I use my fix bug command for minor issues (new chats - 1 per bug/change)
document command - if needed (new chat)
commit command - uses git diff for change summary etc (new chat)
Everything works very cleanly and reliably because I give every chat very specific context and tasks to complete.
I also have other commands for large refactors, audits, setup etc. And all new projects start from my template that I created so I don’t have to redo a bunch of setup steps.
AI is only as good as the context and oversight you give it.
Also I never use ask or plan modes since I have commands for that and they strictly prohibit changes etc.
@mitch_dev hey thanks for this. I never used commands before. Spent a bit of time learning.. seems very powerful, I will need to incorporate that into my daily tools.
Thanks!
Still doesn’t account for 2.0 agent being a lot lazy though. I think Cursor switched to Composer - their model, and you can tell that it is a step backwards. Do you know what they were using before 2.0 on Auto mode?
Hey, auto mode chooses one of the premium models, including Claude 4.5, GPT 5 and others, depending on what is needed at the moment. It may happen that some specific models are more lazy than others and they would require additional prompting.
I’ve been tracking what models the auto mode uses over the last few weeks.
I had 2 weeks of almost 100% sonnet 4.5 - even for small simple tasks.
Now codex 5.1 is released it looks to be the new default model and I’m getting it for almost 100% of the auto chats over the past few days.
The models are just prediction engines so if your codebase is messy and unorganized or your giving vague or confusing prompts/context then the models will struggle.
Spend 30% of your time cleaning up the codebase and ensuring it follows consistent patterns and normalization.
Create a list of commands and regularly update them as you learn new improvements. Use AI to update and enhance the commands - it’s very good at it.
If you have a clean codebase and clean context/prompts then you’ll get high quality output. I’ve found both sonnet and codex to be extremely reliable.
I think we should close this because since the release the agent auto mode has improved. I don’t know exactly what is going on behind the scenes with auto, but what I imagine is that the agent is picking models for what it things is appropriate for the task and I think this is where it doesn’t always shine. There are days when auto is great and days where I could literally throw it out of the window. Maybe that’s a clue. I also started using commands thanks to @mitch_dev’s recommendation which I believe does help.
Anyway, thanks again for all your hard efforts.. it’s a truly remarkable tool.