> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cohesive.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Control agent permissions

> Choose when commands require confirmation and review actions before they run.

Use **Permission mode** in the composer to choose when the agent asks before running commands.

| Mode | Behavior |
| - | - |
| **Manual** | Commands that change state ask for confirmation. |
| **Auto** | The agent decides how to proceed; risky commands still ask. This is the default. |
| **Allow all** | Commands are automatically accepted for this session. |

These modes control command confirmations. They don't grant access to files, Canvases, or connected accounts you aren't authorized to use.

## Change the mode

<Steps>
  <Step title="Open the permission control">
    Click the permission icon in the chat composer. Its tooltip shows the current mode.
  </Step>

  <Step title="Choose a mode">
    Select **Manual**, **Auto**, or **Allow all**.
  </Step>

  <Step title="Send your next message">
    The choice takes effect with that message. Changing the picker doesn't approve a confirmation that's already waiting.
  </Step>
</Steps>

## Review a requested action

Read the command and its inputs before approving it. For a message, check the destination and wording. For a record update, check the record and fields. If the proposed action is wrong, decline it and tell the agent what to change.

When you need to review wording first, ask for a draft or a proposed update explicitly. See the individual [Connector pages](/product/connectors/overview#available-connectors) for supported actions.

<Warning>
  **Allow all** skips command confirmations in that session. Use it only for work you're ready to let run without reviewing each command. The session displays a notice while this mode is active.
</Warning>

You can restore confirmations from that notice, or choose another mode for your next message. See [Sessions](/product/agent/sessions) to start a separate conversation with different instructions.