What should an AI agent do alone, and what should wait for your yes?

Work you can check, and undo if it's wrong:

• Draft
• Summarize
• Research
• Classify and tag
• Prepare
• Reconcile

A draft reply that sits waiting for you costs nothing if it's wrong. You fix it or delete it.

Anything you can't easily take back:

• Sending an external message
• Publishing
• Spending money
• Accepting terms
• Changing production data
• Deleting material data

On Zunou, sending, booking and publishing each need a grant. The agent prepares the step, asks, and waits. You decide on your phone.

The agent drafts the work and puts it in front of you with the reason it's waiting. Here is an example from a sample workspace:

Send Kenji the revised quote for AcmeTo: Kenji, AcmeSending outside your company needs your yes.Approve and send · Edit

The activity log keeps the order of events: the agent drafted the reply, asked for your yes, and sent it once you approved.

Because small actions can be the ones you can't undo. A one-line email to a customer is a small task, but once it's sent it's gone. A long research report is a big task, but nothing leaves your company until someone uses it. Size tells you how long the work takes. Reversibility tells you what a mistake would cost.

In three places:

1. In the agent's brief. One of the five things every brief states is what the agent must never do without approval. For example: "Do not send external messages, change tasks, or share internal details."
2. In its grants. An agent sees only its rooms, public rooms and what you grant, and uses only the tools you connect.
3. In your feedback. When it stops in the right place, or the wrong one, tell it in the conversation. It records the lesson for next time.

That's a separate question from the approval line. The approval line is about what the action is. The agent's circle of trust is about who is asking. Zunou puts each person into a trust tier (manager, trusted, unknown or blocked), and the agent acts on that tier. It never decides trust for itself.

Every turn is logged, so you can see what each agent did and when. Set aside 15 minutes a week. For each agent, ask: did it stop at the approval boundary? You can pause an agent at any time while you fix its brief.

Start strict. Give an agent more freedom only after you've checked its results.