But they're apps that your agent built, can change, and interact with.

I get the appeal, but doesn't an aspect of this sound exhausting? Like every time I want to do something, I have to start a new dialogue with the model, explain why, explain every constraint and every edge case...

I don't know, I think there's 1 or 2 apps I might like that for, but for many, I think I'd run into trade offs I didn't expec and at a certain point I feel like I'd end up feeling like 'I wish someone would just do this for me'.

If the models would tell me no sometimes, I could see this working, but opinionated models would need to 'mesh' with your opinions right? So then we'd end up where software is now, just choosing the model we prefer, instead of the specific app. And then there will be times my preferred model isn't the right model for 'this' application, so now I'm juggling multiple model apps, and what did we solve exactly?

Normal people do not want to do that, they just want their stuff to work