The real expectation shift should be around ownership. Once a SaaS can take action, I no longer want it to just tell me something needs attention, I expect it to either handle it or clearly explain why it stopped.

I had be interested in the rollback side of it. If a natural-language change improves one part of the agent but unexpectedly breaks another, how quickly can you identify and reverse that change?

One thing I will add is auditability. If an agent makes every process consistent, teams should also be able to see exactly what it did, why it did it, and where it deviated, especially when something goes wrong.