People need to be responsible. They are all lazy. Happy to give away any agency and responsibility to a computer program and scream bloody murder when it goes awry.
People give the keys to their kingdom to some program that will even hallucinate and do all sorts of stuff. WHY. it's stupid. that's not harsh, that's the truth.
Before AI (LLMs), people were much more careful with this sort of stuff, but now they give their keys away to some program which has tons of agency and expect that program to somehow be responsible on their behalf. it's stupid.
Be responsible. LLMs even if tailored truly correctly with all the safety guards can just hallucinate and do shit anyway, so if you don't lock them down, or if you give them your secrets, that _will_ end badly at some point.
Even if hallucinations only happen 0.0001% of the time, that is 1 in 10000, and you do tons and tons of tool calls and model calls even if you only put in 1 prompt, so really, the probability is HIGH.
Sure, the probability of that hallucination also involving a secret is much lower, but it is still HIGH if you consider using LLMs for months or years...
Choosing how to sub-divide abilities, and allowing the user to manage that, will be an art unto itself, and depend largely on the system. Thinking a banking API with a rule like "Allow sub user to send payment to water utility once per month, not to exceed $200". The user only needs to intervene if something goes wrong and the amount goes over.