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

# Permission rules

> Say once what your AI coworker may always do, must always ask about, and may never do — instead of approving the same thing every day.

Approving the same action every day is a waste of your attention. Permission rules let you decide once.

## The three answers

Every rule gives one action one of three answers.

|                  | What it means                     |
| ---------------- | --------------------------------- |
| **Always allow** | Go ahead, don't ask me            |
| **Always ask**   | Stop and check with me every time |
| **Never**        | Don't do this, at all             |

Anything you haven't written a rule for keeps the default:

> Actions no rule matches keep the default: sends, changes, and deletions pause for your approval.

So you're not starting from nothing. Rules are adjustments to a sensible baseline, in either direction.

## When you'd write one

**To stop being asked** about something routine that's obviously fine. Your coworker posting to the team channel it always posts to.

**To be asked** about something the default lets through. A coworker allowed to update records, where you'd rather see the ones touching pricing.

**To rule something out entirely.** Deleting. Emailing outside the company. A specific app you attached for reading only.

<Tip>
  **Start with the "never" rules.** They're the ones with real consequences and they take two minutes. Loosening things to reduce prompts can wait until you know what you're actually being asked about.
</Tip>

## Where they live

On the coworker's **Files** tab, in the file labelled **Permissions**.

Rules take effect from your coworker's next step — you don't need to restart the task.

<Note>
  A rule can take power away, and it can remove a prompt for something your coworker was already allowed to do. It cannot grant a capability that's switched off. If [Talking to people](/cowork/capabilities/talking-to-people) is off, no rule makes it possible.
</Note>

<Warning>
  If the permissions file can't be understood, the whole thing is rejected rather than partly applied. Your coworker falls back to the default — asking before sends, changes and deletions — instead of running on a half-read rule set.
</Warning>

## The strictest rule always wins

When more than one rule matches the same action, the **strictest** of them applies:

<Frame caption="Never beats ask, and ask beats allow — whatever order the rules are in.">
  <img src="https://mintcdn.com/glorium/JBtd_griZWPnHu68/images/cowork/control/permission-rules/01-precedence.webp?fit=max&auto=format&n=JBtd_griZWPnHu68&q=85&s=99d5181378ab5e0fbbddfd097a79eca1" alt="A precedence diagram showing never beating ask, and ask beating allow" width="1190" height="146" data-path="images/cowork/control/permission-rules/01-precedence.webp" />
</Frame>

Two consequences worth knowing:

* **Order doesn't matter.** Moving rules around can never weaken your policy, so you can't accidentally undo a "never" by adding something below it.
* **A narrow rule can't loosen a broad one.** If you've said never delete anything, adding "always allow deleting drafts" does *not* create an exception — the never still wins. To carve out an exception, narrow the original rule instead of layering another on top.

## A sensible starting set

For a coworker with real access to your systems:

* **Never** delete anything
* **Never** email outside the company
* **Always ask** before anything touching pricing or contracts
* **Always allow** posting to its own team channel
* **Always allow** reading from any connected app

That takes a few minutes and removes most of the anxiety about giving a coworker real access.

## Rules, connections, or capabilities?

Three levels of control, from broadest to finest.

|                                                                  | Controls                | Use when                                         |
| ---------------------------------------------------------------- | ----------------------- | ------------------------------------------------ |
| **[Capabilities](/cowork/capabilities/overview)**                | Whole abilities         | The coworker never needs this at all             |
| **[App connections](/cowork/capabilities/connecting-your-apps)** | Which actions in an app | Read yes, write no                               |
| **Permission rules**                                             | What needs asking about | The action is fine, the *asking* is the question |

Reach for the finest one that solves your problem. Switching a capability off to avoid a prompt is a blunt fix that costs you the capability.

## Common questions

<AccordionGroup>
  <Accordion title="Do I need rules to be safe?" icon="shield-halved">
    No. The default already pauses for sends, changes and deletions. Rules are for tuning, not for basic safety.
  </Accordion>

  <Accordion title="Can a coworker change its own rules?" icon="lock">
    It can propose a change, and it lands on the Improvements tab for your approval like anything else. It can't quietly widen its own permissions.
  </Accordion>

  <Accordion title="Do rules apply to every task?" icon="repeat">
    Yes — they belong to the coworker, so every task it runs uses them.
  </Accordion>

  <Accordion title="What if two rules could both apply?" icon="code-branch">
    The **strictest** one wins — never, and never asks, both beat allow. Being more specific doesn't override it, and neither does being further down the list. See "The strictest rule always wins" above.
  </Accordion>

  <Accordion title="Can different coworkers have different rules?" icon="users">
    Yes, and they should. A research coworker and one with access to your CRM warrant very different rules.
  </Accordion>
</AccordionGroup>

## Next steps

<CardGroup cols={2}>
  <Card title="Approvals" icon="hand" href="/cowork/control/approvals">
    What asking looks like.
  </Card>

  <Card title="Connecting your apps" icon="plug" href="/cowork/capabilities/connecting-your-apps">
    Control at the action level.
  </Card>

  <Card title="Capabilities overview" icon="toolbox" href="/cowork/capabilities/overview">
    The broadest control.
  </Card>

  <Card title="Automatic checks" icon="clipboard-check" href="/cowork/control/automatic-checks">
    Rules that act rather than ask.
  </Card>
</CardGroup>
