Every agent you deploy sits somewhere on a dial. At one end, it asks a human before doing anything — safe, but barely faster than doing it yourself. At the other, it acts freely — fast, until the day it does something expensive that nobody caught.
Neither end is where you want to live. The interesting work is figuring out, task by task, exactly how far to turn the dial.
Two ways to get it wrong
Too many approvals. Every action routes through a person. The "automation" becomes a queue of little permission requests, and your team spends the day clicking approve. You've added a step, not removed one — and people quickly start rubber-stamping, which is worse than no approval at all because now the check is theater.
Too few. The agent acts on everything. Most of the time that's fine, which is exactly what makes the failure dangerous: nobody's watching when it refunds the wrong customer, emails the wrong list, or books something that can't be un-booked.
The question isn't "should a human be in the loop." It's "in the loop for which decisions" — and the answer is almost never all of them or none of them.
Sort actions by two things
To set the dial well, score each action an agent can take on two axes.
- Reversibility — if it's wrong, how hard is it to undo? Drafting a document is trivially reversible. Sending money isn't.
- Blast radius — if it's wrong, how far does the damage reach? A note on an internal record affects one record. An email to your whole customer list affects everyone at once.
Plot actions against those two and a clear map appears:
- Low stakes, reversible → let the agent act freely. Drafting, tagging, looking things up, preparing work for review. Gating these just adds friction.
- High stakes or irreversible → require an approval before it fires. Anything that spends money, leaves the building, or can't be taken back.
- In between → let it act, but log everything and make it easy to catch and reverse after the fact.
Approval should be one clear decision
When an approval is warranted, its design matters as much as its existence. A good approval is a single, legible decision: here's what the agent wants to do, here's why, approve or edit or deny. A bad one is a vague notification that piles up until someone bulk-approves the backlog without reading it.
The pattern that holds up is draft-then-approve: the agent does all the work — writes the email, prepares the refund, fills the form — and stops at the last step, presenting a finished action for a human to release or adjust. The person supplies judgment, not labor. (It's the same principle behind agents that finish the job rather than the sentence.)
The dial should move
The right setting isn't fixed. A new agent starts cautious — more gated, heavily logged — because you haven't earned trust in it yet. As it proves itself on real work and the traces show it making the right call, you turn the dial up: fewer approvals, more autonomy, on the actions it's demonstrated it can handle.
That progression only works if you can see what the agent has been doing — which is why observability isn't separate from autonomy. You earn the right to loosen the dial by watching. We treat "when should an agent act, ask, or defer" as a design problem in its own right; it's one of the standing programs in Nexos Research.
If you're weighing how much to let an agent run on its own, that's a conversation worth having before you deploy, not after. Book a free audit and we'll map your workflow's actions onto the dial with you.