How much do you trust AI agents with internal systems?

Auditability, for me — but slightly upstream of just logging what was checked. The agents I've ended up trusting are the ones that state their own confidence level explicitly (this is settled vs. this is still provisional) rather than reporting every result the same flat way. A good log still matters, but it only helps once you know which outputs were worth auditing in the first place.

Auditability, and it tends to get picked last. An agent that fails loudly is a nuisance you fix in an afternoon. One that reports success while having done nothing is the expensive kind, and you only catch it if the log records what was actually checked rather than what the agent believed it checked.