Hey guys, I’m a developer who uses Cursor every single day for hours. When you first made it mandatory to use the agent and removed the chat feature, I came here and wrote some huge complaint posts — and people were supporting those posts, so I definitely wasn’t alone in feeling that way. I’d usually come here just to complain, but I wanted to share some positive feedback this time.
I recently switched to the agent at version 0.51 due to decommisioning of version 0.45 and I’ve noticed that it’s significantly improved. So, keep up the good work, guys! I just wanted to say thanks and encourage you to keep developing it further. Great job! Love Cursor
1 Like
Hey, I’m sorry, but I need to take back what I said earlier. The way the agent works is so unpredictable, making even simple tasks unnecessarily complex. Unfortunately, it’s just bad architecture. I’m currently trying to fix the broken dataset it created for my DWH. Please always keep the manual mode available, I just can’t trust the agent at this point.
The agent is great for a lot of uses, but there are situations where it’s not quite fit for what you are trying to do!
I’d love to know a bit more about what went wrong with your use of Agent, as we’d love to improve on wherever it went wrong for you!
I gave a perfectly working piece of code and just wanted to make a small change to enrich a column, but it ended up completely messing up my ETL process. It was working fine and keeping the data clean before. The change I asked for had nothing to do with the data load part, but instead, it went totally off the rails, freestyling and wrecking the ETL. Now I’m stuck for hours cleaning up the duplicate records it created. About Agent, I liked being able to run code directly from the terminal or Composer and run SQL queries from there. But coding-wise, this has been a mess,the agent is hallucinating so badly it made a whole working codebase useless. Making me lose hours of time.
This can really depend on what model and setup you are using unfortunately, as even the better models can be uncontexualized and make bad decisions.
In this case, I’d recommend reviewing the code it writes and also using checkpoints to roll back your code to an earlier point, if Cursor made a decision you didn’t catch!
o3 is now much cheaper and is a very good model for learning everything it needs before making any changes, so may be a good model for this case!
Thanks for the feedback, but if I need to review the code on every single iteration with the agent, it kind of defeats the purpose of using an agent in the first place. I’m going to stick with manual mode since it’s much more precise for my needs.
The agent is great for vibe coders with less architectural knowledge or less dev experience, but for experienced devs, manual mode is just a better option.
If there’s a way to run queries in manual mode without having to jump back and forth between the terminal, that would be perfect.
Also, please, please, please, don’t remove manual mode!
You can use a custom mode for this! You can disable all the read tools, which forces the model to only see what files you @ in as context.
However, you can also enable the terminal tool so it can do exactly this!
You can learn more about custom modes here:
1 Like
oh wow, looking into this, thanks for sharing team cursor 
1 Like
Custom mode is what i was looking for, thanks for sharing!!
