[margorczynski]

Did you ever you the newest LLMs with a harness? Because I usually hear this kind of talk from people whose most recent interaction was with GPT-4o copy-pasting code into the chat window.

Maybe I'm biased but I don't buy someone truly thinking that "it's just a tool like a linter" after using it on non-trivial stuff.

[overgard]

I'm using Claude Code (and Codex) (with the expensive subscriptions) on an app I'm building right now. I'm trying to be maximalist with them (to learn the most I can about them .. and also that subscription isn't cheap!). My impression, and yes, this is using the latest models and harness and all that would agree with the GP. They're a very handy tool. They make me faster. They also do a lot of things that, as a professional software developer, I have to frequently correct. They duplicate code like nobodies business. They decide on weird boundaries for functions and parameters. They undo bug fixes they just made. I think they're useful, but the hype is out of control. I would not trust software made with these tools by someone that couldn't write that software by hand. It might work superficially, but I'm definitely not giving any personal data to a vibe coded app with all the security implications.

[qudat]

I use it pretty extensively. The reason why it's a tool is because it cannot work without an SWE running it. You have to prompt it and re-prompt it. We are doing a lot of the heavy lifting with code agents that people hyping it are ignoring. Sure, as a non-swe, you can vibe a project from zero-to-proto, but that's not going to happen in an enterprise environment, certainly not without extensive QA/Code review.

Just take a look at the openclaw codebase and tell me you want to maintain that 500k loc project in the long-term. I predict that project will be dead within 6 months.