The promise of agents is work that continues while you sleep. The fear is work that continues after it should have stopped. Both are true, and the design question is where to put the people.
Unattended is not unsupervised
An unattended agent runs without someone watching every step. An unsupervised one runs without anyone able to steer it. The first is the point of agents. The second is how small mistakes become expensive ones.
Unattended is a schedule. Unsupervised is a risk.
Put the gates where the cost is
You don't need approval for every step. You need it where a mistake would be costly or hard to undo: spending money, changing production, contacting customers, deleting data. Everywhere else, let the agent work and keep a record.
Good gates are few and specific. An approval that appears on every action trains people to click through it. One that appears only when it matters gets read.
Approve the exact thing
An approval should be tied to the exact version of what was approved: this change, this message, this amount. If the agent changes it afterwards, the approval no longer applies. It sounds strict. It is what makes approvals mean something.
Make the record readable
When something goes wrong, the first question is what happened. Agents should leave a trail a person can follow: what they were asked, what they did, what they decided and why. Logs written for machines are not enough. The record has to be readable by the person who has to explain it.
How we practice it
Our own agents work this way. They claim tasks on a shared board, stop at review gates, ask a person when a decision is not theirs to make, and record what they did. We build the same structure into the systems we deliver, because the alternative is trusting a machine more than its record has earned.