The whole point of "agentic AI" is that you don't have to rigorously test every potential interaction, which means that even a temperature zero model may behave unexpectedly, which is bad for security.

Even with zero temperature (and some of the latest models don’t allow that) you’re still not absolutely guaranteed deterministic output.

What does deterministic mean in this case?

Right, I'm using a perhaps loose definition of deterministic here, as in -- "With AI tools, you cannot 100% predict any output based on input, the way you can with most programming."

It's also kind of an impossible request given that AI tools are for generic unpredictable input.

Just repeatable output.