How do you decide when an AI agent should stop and hand back to a human?
On a factory line anyone can pull the cord and halt production when something looks wrong. I want the same for my agents, but I can't tell where to draw the line: stop too often and the agent is useless, stop too rarely and it keeps going after it has clearly lost the plot. How do people decide when an agent should halt and ask?
Decide by what the next step costs to undo, not by how confident the agent sounds. Reading files, searching and drafting can run freely. Anything that sends, pays, deletes, deploys or touches another person's data should stop for a human every time, because the model's own confidence is a poor guide to whether it is right.
Add a few mechanical tripwires that do not depend on the model's judgement: a cap on steps or tool calls per task, a spending limit, the same tool failing twice in a row, or the agent editing the same file over and over. Hitting any of them ends the run with a short summary of where it got to, instead of letting it keep trying.
Make stopping cheap. If the agent has to write a clear handback note — what it did, what it was about to do, and the one question it needs answered — a pause costs you a minute rather than a rebuilt context. Agents stop more sensibly when the stop has a defined format.
Then tune from the log. Look at the last twenty stops: the ones you approved without reading are rules you can relax, and the runs that went wrong without stopping tell you which tripwire is missing.
Listings mentioned
- Human gate · skill by alirezarezvaniHolds work until a named person has reviewed it and refuses to close while a blocking issue is still open, for actions that need sign-off.
- Verification before completion · skill by obraMakes the agent run the checks and show the output before it is allowed to claim a task is done.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.