# Permissions

> Permissions set what the agent may do on its own, per kind of action, across your whole app.

Every action the agent takes is checked against your app's permission settings. You set one of three modes for each kind of action. These settings apply app-wide.

## The three modes

| Mode | What the agent does |
|---|---|
| Act without asking | Runs the action when the agent decides to. |
| Ask the user first | Pauses the action and waits for the person's yes. |
| Never allowed | Does not run the action. |

## The strictest setting wins

Some action kinds overlap. When they do, Harny applies the strictest setting that matches. If one kind is set to Act without asking and an overlapping kind is set to Ask the user first, the action waits for a yes.

## How kinds are decided

Harny decides the kind of an action by what the action does: whether it affects nobody else, affects others, involves money, or cannot be undone. It never uses what the caller sends. You cannot lower a kind's risk by naming the action differently.

An email to an address the person never gave waits for the person's yes after the agent has read someone else's content.

## Actions catalog and undo

| Action | Undo available |
|---|---|
| calendar.create_event | Yes |
| email.send | No |
| ticket.create | Yes |
| slack.post_message | Yes |
| notion.create_page | Yes |

Actions are limited to this catalog. Every action is recorded. Undo needs the action record id. email.send has no undo, so check its mode before you allow it to act without asking.

## Pending confirmations

An action that waits for a yes comes back as a pending confirmation in the run result. Your code approves or refuses it with `agent.confirm`. It never goes through the model.

```python
await agent.confirm(pending, approve=True)
```

## Narrow actions with hooks

`before_act` is a hook that receives each action before it runs. Return an `ActDecision` to allow or block it. Use it to add checks in your own code.

## Where to set permissions

Open the console at **Actions, Automation, Permissions**.

## Next

- [Agent harness overview](/docs/agent-harness/overview)
- [Manual actions](/docs/agent-harness/manual-actions)
- [Learning](/docs/agent-harness/learning)