Ask HN: Cursor or Windsurf?

I’ve been using Zed Agent with GitHub Copilot’s models, but with GitHub planning to limit usage, I’m exploring alternatives.

Now I'm testing Claude Code’s $100 Max plan. It feels like magic - editing code and fixing compile errors until it builds. The downside is I’m reviewing the code a lot less since I just let the agent run.

So far, I’ve only tried it on vibe coding game development, where every model I’ve tested struggles. It says “I rewrote X to be more robust and fixed the bug you mentioned,” yet the bug still remains.

I suspect it will work better for backend web development I do for work: write a failing unit test, then ask the agent to implement the feature and make the test pass.

Also, give Zed’s Edit Predictions a try. When refactoring, I often just keep hitting Tab to accept suggestions throughout the file.

Can you say more to reconcile "It feels like magic" with "every model I’ve tested struggles."?

It feels like magic when it works and it at least gets the code to compile. Other models* would usually return a broken code. Specially when using a new release of a library. All the models use the old function signatures, but Claud Code then sees compile error and fixes it.

Compared to Zed Agent, Claude Code is: - Better at editing files. Zed would sometimes return the file content in the chatbox instead of updating it. Zed Agent also inserted a new function in the middle of the existing function. - Better at running tests/compiling. Zed struggled with nix environment and I don't remember it going to the update code -> run code -> update code feedback loop.

With this you can leave Claude Code alone for a few minutes, check back and give additional instructions. With Zed Agent it was more of a constantly monitoring / copy pasting and manually verifying everything.

*I haven't tested many of the other tools mentioned here, this is mostly my experience with Zed and copy/pasting code to AI.

I plan to test other tools when my Claude Code subscription expires next month.

Zed's agentic editing with Claude 3.7 + thinking does what you're describing testing out with the $100 Claude Code tool. Why leave the Zed editor and pay more to do something you can run for free/cheap within it instead?