Human-in-the-Loop

A design pattern where the agent pauses at defined checkpoints and waits for a human to review or approve before continuing — rather than running fully autonomously.

Example: An email-drafting agent writes the message and shows it to the user before sending. The user clicks 'approve' or 'revise'.

In practice: The right level of human involvement depends on the stakes. Irreversible actions (sending money, deleting data) almost always warrant a human checkpoint. Low-stakes retrieval tasks usually don't.

• is a form of GuardrailsRequiring human approval at key steps is a behavioral guardrail, implemented at the workflow level rather than as a filter.