Ask HN: Do you use LLMs to perform tasks in real applications?

I'm referring to some sort of deterministic tasks (like extract something from a user input, and probably any function call). If so, do you test them? How? Are you concerned about it stop working with every prompt change, or even underlying model changes? Full disclosure: I'm working on a solution in this space.