[safety1st]

I'm not at all convinced that "break your concentration and go check on an agent once every several minutes" is a productivity increaser. We already know that compulsively checking your inbox while you try to code makes your output worse. Both kill your focus and that focus isn't optional when you're doing cognitively taxing work--you know, the stuff an AI can't do. So at the moment it's like we're lobotomizing ourselves in order to babysit a robot that's dumber than we are.

That said I don't dispute the value of agents but I haven't really figured out what the right workflow is. I think the AI either needs to be really fast if it's going to help me with my main task, so that it doesn't mess up my state of flow/concentration, or it needs to be something I set and forget for long periods of time. For the latter maybe the "AIs submitting PRs" approach will ultimately be the right way to go but I have yet to come across an agent whose output doesn't require quite a lot of planning, back and forth, and code review. I'm still thinking in the long run the main enduring value may be that these LLMs are a "conversational UI" to something, not that they're going to be like little mini-employees.