Why agents are bad pair programmers

I partly agree, and I think to some extent the reaction people have to LLM-assisted programming is because they produce so much code. It’s difficult to process that much so quickly, especially if you’re looking for mistakes or subtle bugs. The reason I only partly agree is that I think IDE-based tools are still useful. I’ve been playing around with Zed’s agent mode and I think their UI for reviewing changes feels so much like a pull request I haven’t needed to use the async agents, and I can work with smaller scale things and accept/reject basically by line.

This post made me dream of an interface where the LLM, instead of dumping code at you, basically edits it with you - line by line or chunk by chunk - with an explanation so you can proceed or disagree for each chunk, and then its response can change as you accept or push back on its changes.

This is also why I have disliked my experience so far, although to be fair I haven’t used some of the more state of the art tools.