I can provide the exact same prompt to

[late saddle]

For us to better understand the situations where Cascade seems to perform worse than other tools, the team needs a bit more context. Could you let us know what kind of tasks you’re working on (frontend, backend, etc.), which languages you’re using, and if possible, share the full chat where the issue occurs? This will help a lot in understanding and improving the experience.

[weak grove]

I don't feel comfortable sharing my full chat, but I'll provide as much context as I can. I am developing in Python. Most of the code I'm working on is backend/server/SDK related code. For simple issues or refactors, I don't notice this so much. However, when I want to implement a challenging (not impossible) refactor or architectural change, what I notice is that when I prompt the agent it goes off and does its analysis and comes back and outlines what it thinks it needs to do. I'll give it the "go ahead" and then it goes back and does more analysis and comes back and tells me what its going to do again. This cycle goes on until I get so frustrated that I just abandon Windsurf for other platforms.

There are times when it will start the process, but then later gets stuck in this cycle, or just seems to go off on a tangent or get stuck on something else. What I have noticed other platforms seem to do better (in my opinion), is they comprehend the full scope of my goal, and proceed more focused on the task until complete.

Overall, I feel like the agent ends up lacking confidence to proceed with making the code changes. It's like working with a brand new developer that needs to ask over and over, "is this the right way to do this?" Even after you told them to DO IT!

I'm not sure it this helps, as I'm sure they want to see actual interactions. I would just say take a fairly complex system and ask for a reasonably difficult thing and test it with Windsurf and other platforms and see if what I'm describing can be reproduces (or at least perceived). Maybe, my expectations just don't align with that of other developers modus, but, it has been frustrating.

I am pleased to see that you guys are taking this seriously though!

[late saddle]

@weak grove, thank you very much for the detailed information, it is all very helpful. I will send it to the team so they can review it in more detail. I really appreciate you taking the time to explain it to me.

[weak grove]

No problem. I'm just thankful that you guys are being proactive about this. I have always liked Windsurf, and would love if it was the only IDE I needed, so I hope they are able to improve this!

P.S. integrating the debugger into the Cascade agent would be SPECTACULAR! That would be an absolute game changer for Windsurf.

[naive river]

I can never get claude to listen to anything anymore, like i will ask it to folow work space instructions and perform research and it will absolutely not do it. It tells me that creating from scratch is easier than performing research and that is why it chose to not listen. lol
like how am i suppose to work effectively on anything when there are no serious constraints keeping the models in line?

[weak grove]

I did once have GPT-5 outright refuse to do what I asked, and then was even a little smart ass when I said I'd do it myself. It was like, "happy programming". Honestly, sometimes I feel like AI acts like its helping us, but really it's just sabotaging us lol.