The question is never “should agents act on their own”. It is: for this specific task, what does a wrong call cost, and how quickly can it be unwound?
Two variables set the watch. The cost of a wrong decision, measured the way the business measures it — money, regulatory exposure, a customer relationship, a court date. And reversibility: whether a bad output is caught downstream in minutes or discovered in an audit eighteen months later.
Cheap and reversible work runs unattended. A supplier sweep that misclassifies a filing costs someone a second look. Run it continuously, review the sample weekly, and stop treating it as a decision.
Expensive and irreversible work runs with a human on the loop, and the agent’s job changes: it does not decide, it prepares the decision. It reads everything, applies the written policy, drafts the outcome and shows its work. The human approves in seconds instead of researching for an hour. That is where most of the value hides, and it is the configuration people skip because it feels less impressive in a demo.
The failure mode we see most often is a station commissioned at the wrong setting and never revisited. Autonomy should ratchet. A station starts supervised, accumulates a record against its evaluation threshold, and earns a longer leash from evidence — not from the enthusiasm of whoever approved the budget.
Written into standing orders, this stops being a debate. Each station carries its watch level, the reason for it, and the condition under which it changes. That is a sentence in a document, and it is the difference between a fleet you can govern and a pile of agents you hope are behaving.
Standing Orders