Human in the loop
Autonomy is a setting, not a switch
A system does not have to be fully manual or fully automatic. You choose, action by action, how much it may do on its own.
Guardrails are the rules around the model: what it may access, what needs approval, and what is never allowed. They are enforced in code.
1. Sort actions by risk
Search documents or look up an order.
Runs automatically.Draft an email or update a record.
A person approves first.Send money or delete data.
Needs approval, or is not allowed at all.2. Make approval easy
Show the person exactly what will happen: the action, its arguments, and why the model chose it. A clear approval screen is faster and safer than reading a long transcript.
3. Escalate when unsure
Give the system a way to ask for help: when a request is ambiguous, when checks fail, or when a limit is reached. Asking is a valid result, not a failure.
4. Learn from corrections
When people reject or edit a proposed action, record it. These cases are your best new evaluation examples.
When is it useful?
Whenever actions affect money, data, or other people, and whenever mistakes are costly or hard to undo. As the system proves reliable, you can move actions to a more automatic level.