08.03 · Concept
Human in the loop, where it counts
Put an approval gate on the actions that are expensive to undo.
Human approval belongs at the tool or permission boundary where an agent is about to do something costly to reverse. Safe agent design lets low-risk work run automatically, but pauses before data deletion, payments, customer contact, permission changes, shell execution, or production impact so a person can inspect and authorise the proposed action.
What this lesson answers
- where should agent approval gates be placed
- which AI agent actions need human approval
- how to design human in the loop agents
Notes
Human in the loop is not asking a person to babysit every model response. It means deliberately stopping automation at the boundary where a bad action becomes costly, slow, risky, or impossible to reverse. Reading a file, drafting a message, or searching documentation can usually proceed automatically; deleting data, sending money, emailing customers, changing permissions, or executing shell commands should often require approval.
A useful mental model is a production deployment pipeline. Tests and builds run unattended, but promotion to production may need a reviewer because rollback has real…
Common questions
- Does human in the loop mean reviewing every agent response?
- No. It means interrupting automation only where a wrong action has serious consequences. Reading, searching, summarising, or drafting can often proceed without review. Actions that affect money, customers, credentials, production systems, permissions, or durable data should be stopped for explicit approval before execution.
- Why put the approval gate around tools instead of prompts?
- A prompt can ask the model to be careful, but it is not an enforcement boundary. The runtime should pause before the tool call is executed, show the proposed action and arguments, and continue only after approval. That makes the control auditable and independent of model behaviour.
- What should an approval request show to the reviewer?
- It should show the exact operation, the target resource, the arguments, the reason the agent wants to do it, and the effect of approving or rejecting it. The reviewer needs enough context to decide quickly without reconstructing the agent’s state from logs or chat history.
Short definition: what is Human in the loop, where it counts?