AI Explained
Human-in-the-loop for AI agents
Would you let an AI agent refund customers on its own? Human-in-the-loop pauses the agent before risky actions so a person can approve, edit or reject.
Would you let an AI agent refund customers on its own? Probably not. But you'd happily let it search orders and draft replies. Human-in-the-loop is how you get both: the agent runs freely where it's safe, and pauses for a person where it counts.
How it works
You sort your agent's tools into two groups:
- Safe tools run on their own: search, read an order, look up a policy.
- Risky tools are marked as needs approval: refund, delete, send email, run a shell command.
When the agent decides to call a risky tool, the run doesn't execute it. It stops, saves its state, and shows a person what it wants to do, for example "Refund $49 on order #1042".
The person then has three choices:
- Approve it as is.
- Edit it, for example change the amount.
- Reject it, optionally with a reason the agent can use.
On approval, the agent picks up exactly where it paused. It doesn't start over.
It's built into the major frameworks
The OpenAI Agents SDK lets you mark tools with needs_approval. When one is called, "the run records an approval interruption instead of executing the tool", and your app resumes the same run after a decision. OpenAI recommends this before side effects like cancellations, edits, shell commands and sensitive MCP actions.
LangGraph offers the same idea with interrupts: pause the graph at a point, wait for human input, then resume with an approve or reject, or with edited tool arguments.
What deserves approval?
A useful test: can this be undone, and how much does it cost if it's wrong?
| Action | Approval? |
|---|---|
| Search, read, summarize | No |
| Draft a reply (not sent) | No |
| Send an email to a customer | Yes |
| Refund, payment, purchase | Yes |
| Delete data, run shell commands | Yes |
Too many approvals and people start clicking "yes" without reading. Keep them for the actions that really matter.
Fast where it's safe. Human where it counts.