Human in the loop means an AI agent pauses before a risky action and waits for a person to approve or reject it. It is the last check on actions that are costly or cannot be undone.
How it works
The agent proposes a tool call. A check on the call decides whether it needs approval. If so, the call is held and a person sees the tool and its arguments. They approve or reject, and the agent continues with the result or the refusal.
Which actions deserve it
Approval on every step makes the agent useless and teaches people to click yes. Keep it for actions that are:
- Irreversible. Deleting data, sending a message.
- Costly. Payments and refunds above an amount.
- Far-reaching. Changes to permissions, production settings or many records at once.
An agent that takes these with nobody confirming has excessive agency. Reads and actions that are easy to undo pass without approval.
How to avoid rubber-stamping
- Ask rarely. Each needless prompt lowers the attention the next one gets.
- Show the real arguments (the recipient, the amount, the command), not the agent's summary of them.
- Make rejecting as easy as approving, and do not run the action if nobody answers.
- Watch the approval rate. A gate that is approved every time is either not needed or not read.
With White Circle
A policy in Control can hold a tool call for confirmation by a person before it runs.
