This sounds like a benefit in some ways, but also introducing a lot more compute per request than if things are well defined.

I suppose the trade offs may depend on each use case, using ai to handle these entirely, or only intervening if an error is returned and the ai can smartly rectify the problem.

Interesting idea. I’m curious to see where people take this