> ## 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.

# How Cowork works

> The mental model behind AI coworkers — how a task runs, what your coworker can reach, and where it stops to ask you.

Five minutes on how this actually works. Worth reading once — most confusion later traces back to one of the ideas on this page.

## Three things, nested

```mermaid theme={null}
flowchart TB
    WS[Your workspace<br/>shared with your team]
    C1[A coworker<br/>Nora, research]
    C2[A coworker<br/>Sam, assistant]
    T1[A task<br/>private to you]
    T2[A task<br/>private to you]
    T3[A task<br/>private to a colleague]

    WS --> C1
    WS --> C2
    C1 --> T1
    C1 --> T2
    C2 --> T3
```

**Your workspace** holds everything, and your team shares it.

**A coworker** is a colleague you've hired — a name, a job, a personality, a set of things it's allowed to do. It's shared with the workspace, so a colleague can give Nora work too.

**A task** is one job. It has its own conversation and its own files, and it's **private to whoever started it**. Your colleague giving Nora a task doesn't interrupt yours, and can't see it.

## How a task runs

Your coworker doesn't follow a script. It works in a loop, the way a person does:

<Steps>
  <Step title="Work out what's needed">
    Read what you asked for, look at what's there, decide the next useful thing.
  </Step>

  <Step title="Do it">
    Search, open a file, run the numbers, use one of your apps.
  </Step>

  <Step title="Look at what came back">
    Did that work? Does it change the plan? This is the step that makes it different from automation — the result of one step shapes the next.
  </Step>

  <Step title="Repeat, or stop">
    Keep going until the job is done, or stop and ask you something.
  </Step>
</Steps>

Everything you see in the conversation is that loop, made visible.

## What it can reach

Three sources, and knowing which is which explains most "why can't it do X" moments.

|                                                            | What it is                                                                                                         |
| ---------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------ |
| **[Built-in capabilities](/cowork/capabilities/overview)** | Web search, reading pages, code and files, a browser, images, self-scheduling, your automations, talking to people |
| **[Your apps](/cowork/capabilities/connecting-your-apps)** | Gmail, Slack, your CRM — whatever you've connected and allowed                                                     |
| **[Its skills](/cowork/skills/overview)**                  | Procedures you've taught it, for jobs with a right way of being done                                               |

If your coworker says it can't do something, it's almost always a switch on its Capabilities tab.

## Where it stops

Your coworker works freely on anything reversible — reading, searching, analysing, drafting. It stops in two situations:

* **It needs permission.** Something is about to leave your workspace or change your systems. See [Approvals](/cowork/control/approvals).
* **It needs information.** Something only you know, and it would rather ask than guess.

Both pause the task until you answer. Neither times out and proceeds anyway.

## What carries, and what doesn't

The thing people get wrong most often.

|                                                       | Across tasks?             |
| ----------------------------------------------------- | ------------------------- |
| The conversation                                      | No                        |
| Files in a task                                       | No — private to that task |
| Its instructions and personality                      | Yes                       |
| Its skills                                            | Yes                       |
| What it [remembers](/cowork/control/memory) about you | Yes                       |

So a new task isn't starting from nothing. Your coworker still knows who you are and how you like things — it just doesn't have that particular conversation.

## Good to know

<Tip>
  **One job per task.** Related follow-ups belong together; a new subject deserves a new task. It's the single habit that makes everything else work better.
</Tip>

<Note>
  Tasks run in the background. Close the tab, start another one, come back tomorrow — nothing needs you watching.
</Note>

## Common questions

<AccordionGroup>
  <Accordion title="Is a coworker a chatbot?" icon="robot">
    No. A chatbot answers; a coworker acts. It opens the file, runs the numbers, sends the email — and comes back with the finished thing rather than instructions for making it.
  </Accordion>

  <Accordion title="Can two people use the same coworker?" icon="users">
    Yes. Coworkers are shared, tasks are private. Everyone gets their own conversations with the same colleague.
  </Accordion>

  <Accordion title="Does it learn from my colleagues' tasks too?" icon="brain">
    Its memory belongs to the coworker, so what it learns benefits everyone using it. The tasks themselves stay private.
  </Accordion>

  <Accordion title="What if it does the wrong thing?" icon="triangle-exclamation">
    Every step is visible, including exactly what it changed in a file. Correct it in the chat and it remembers.
  </Accordion>
</AccordionGroup>

## Next steps

<CardGroup cols={2}>
  <Card title="Quickstart" icon="rocket" href="/cowork/quickstart">
    Put it into practice.
  </Card>

  <Card title="What a coworker can do" icon="toolbox" href="/cowork/capabilities/overview">
    Every capability, one page each.
  </Card>

  <Card title="Cowork vs workflows" icon="scale-balanced" href="/cowork/vs-workflows">
    When to use something else.
  </Card>

  <Card title="Approvals" icon="hand" href="/cowork/control/approvals">
    Where it stops for you.
  </Card>
</CardGroup>