The key to productivity with these tools is a tight feedback loop.

Cursor has an amazing tab completion model. It uses a bunch of heuristics to try to understand what you're doing and accelerate it. It's a bit like a very advanced macro engine, except I don't have to twist my brain into a pretzel trying to program my editor - it just works, and when it doesn't I press Esc, which takes about half a second. Or I reach for the agentic mode, where the iteration cycle is more like 30 seconds to a minute.

With the fully agentic editors, it takes more like 15 to 30 minutes, and now it's a full on interruption in my flow. Reviewing the agent's output is a chore. It takes so much more mental bandwidth than the quick accept/reject cycle in the editor. And I have to choose between giving it network access (huge security hazard) or keeping it offline (crippling it in most codebases).

I find that for throwaway codebases where I don't care about maintenance/security/reliability, I can put up with this. But for most of my work it produces poor results and ends up being a net negative for productivity - for now.

I'm sure this will improve, but for now I consistently get better results in Cursor.