This entire thing with non-deterministic agents where you can’t be sure what’s going to happen each time has always felt like Sorcerer Mickey and the Mop and the unintended consequences. But the deal with that story is that Mickey really really wanted to get some mopping done and was totally OK with the risk of, I don’t know, drowning everyone until everyone started getting drowned. These agents are stupendously useful to people and people are getting a lot of value out of them. And they’re horrifically insecure. The value and the lack of security and risks come hand in hand. Like I’ve kept saying, to be most useful, they need to know what you know, and to be able to do what you do. The whole deal with an agent is, if I may American High School Essay myself, that a dictionary defines it as “a person who acts on behalf of another person or group”, like, literally a thing that is empowered to do things for you. There have been some pretty shitty human agents! There are whole laws in England and Wales about the duties that agents owe to their clients. As soon as you let something act as or for you, there are risks. The deal with computing and software is that we think we can mitigate those risks. Sure, we can, a bit. Maybe. But probably not as much as we want to, because the world is just too messy with too much ambiguity and need for flexibility. It will always, always, be easier and faster and less hassle to just let something else do it.