Is anyone actually getting 'agent autonomy' right in production, or running supervised workflows?

Autonomous agents' is the marketing line. Production reality looks different.

Almost every production agent I've seen has:

- A human approval gate somewhere in the flow

- A predefined set of allowed actions

- Logging that gets reviewed daily by an ops person

- Hard stops on anything financial or customer-facing

Which is basically supervised automation, not autonomy.

Is anyone actually running truly autonomous agents in production with real consequences attached? Genuinely want to know how that scales without going wrong.