Human in the loop for AI agents: approve risky actions before they run
An AI agent that can use tools can send, share and delete. Human in the loop means a person says yes before the risky ones happen. Done badly it is a flood of pop-ups. Done well it is a few decisions a day.
What human in the loop means for an agent
A chat model only writes text. An agent calls tools: it reads a calendar, edits a document, sends a message. Human in the loop (HITL) puts a checkpoint between the agent's decision and the tool: the call is held, a person reads what was asked, and it runs only after a yes.
The point is not to review everything. It is to review the small set of actions that can hurt: anything that leaves your organisation, anything that cannot be undone, and anything that changes who has access.
Which actions to hold, and which to let run
- Let run: reading, and writing to your own things.
- Hold for approval: sharing outside, sending to someone new, changing settings, adding forwarding or filter rules, bulk actions.
- Hold with stronger proof: what cannot be undone, such as sending a message or deleting. A tap is not enough; ask for a fingerprint or face.
- Refuse outright: anything the agent's role never needs.
Classify by what the action does, not by what the tool's description says. A description is text someone else wrote, and it is the usual place to hide instructions for the agent.
How to avoid approval fatigue
If every call asks, people approve without reading. Three things keep the number down: sensible defaults (reads never ask), rules for what you have already approved ("this agent may share with this person"), and limits that turn a burst into a pause instead of a hundred prompts.
Measure it. Count how many actions ran on their own, how many waited and how many were refused each week. If waiting is most of the total, the rules are too tight; if nothing ever waits, they are too loose.
What a good approval looks like
- The person sees the whole request, exactly as the agent sent it.
- Hidden characters and look-alike addresses are pointed out.
- The approval is for that exact request and runs once. Changed arguments are a new request.
- It expires. A request nobody answers closes, and the agent has to ask again.
- The decision leaves a record that cannot be quietly edited.
How Operon OS Agenaut does it
Agenaut sits between your agents and their tools. Each agent has a level (Observer, Assistant or Trusted). Reads and ordinary writes run; sharing outside, sending to someone new, deleting, settings and forwarding rules wait for you. You approve from your phone or computer, and what cannot be undone asks for your passkey. Every decision gets a signed receipt.
A request waits 15 minutes by default. After you approve the same thing twice, Agenaut offers to make it a rule.
Try it. Operon OS Agenaut is in beta. Request access, or read the docs to see how an agent is connected.
Related guides
- AI agent governance: a practical checklist
- What is an MCP gateway, and when do you need one?
- MCP security: tool poisoning, prompt injection and how to defend an agent
Terms used here are explained in the glossary.