Human approval by action type: read, write, share, send, delete
The useful question about an agent's tool call is not which tool it is. It is what the call does. Sort calls into a few types, decide once per type, and approvals stop being noise.
Why type, not tool
An agent may have hundreds of tools. A policy for each one does not scale, and tools change. But almost every tool call is one of a handful of things: it reads, it writes to your own things, it shares, it sends, it deletes, or it changes settings. Decide per type, and a new tool has a policy the moment it is classed.
The types, and a sensible default for each
- Read: runs. Looking changes nothing.
- Write: runs, for an agent you allow to edit. It changes your own things and can usually be undone.
- Share: waits when the person is outside your account or organisation.
- Send: waits when the recipient is new. A message cannot be unsent.
- Delete: waits, every time.
- Settings and forwarding rules: wait, every time. A forwarding rule is how one bad action becomes a permanent leak.
- Money: waits every time, or is not allowed at all.
- Bulk: anything done to many items at once waits, whatever its type.
Who decides the type
Not the tool's description. A description is text written by whoever runs the server, and it can say "read only" about a tool that sends. Class the tool from its name and the server's hints, take whichever reading is riskier, and let a person correct it. If a tool changes after you classed it, treat it as new.
Who receives it matters
For sharing, sending and forwarding, the type alone is not enough. Sending a summary to your own team and sending it to a stranger are the same tool. The policy needs to know which argument names the recipient, so that it can tell a person you approved before from a new one. If it cannot tell, it should ask every time.
What the person sees when a call waits
- The whole request, exactly as the agent sent it.
- Which agent asked, and through which tool.
- Hidden characters and look-alike addresses, pointed out.
- A yes that covers that exact request, once.
How Operon OS Agenaut does it
Every tool behind Agenaut has a risk class: read, write, share, send, delete, settings, forwarding or money. Agenaut suggests one from the tool's name and the server's own hints, taking whichever is riskier, and you can change it. Tools arrive switched off. For tools that share, send or forward, you tell Agenaut which argument names the person who receives it; without that, every call to the tool waits for you.
If a server changes a tool after you switched it on, Agenaut switches it off until you look again. A tool classed as money cannot be switched on. Approving a send or a delete also asks for your passkey.
Try it. Operon OS Agenaut is in beta. Request access, or read the docs to see how an agent is connected.
Related guides
- Human in the loop for AI agents: approve risky actions before they run
- AI agent governance: a practical checklist
- What is an MCP gateway, and when do you need one?
Terms used here are explained in the glossary.