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

# AI agent: behaviour & instructions

> Decide how your chatbot answers: instructions, tone, model level and creativity.

The **knowledge base** decides *what* your chatbot talks about; the **AI agent**
decides *how* it does that.

## Where to find it

<Steps>
  <Step title="Open your chatbot">
    From **Chatbots**, click the chatbot you want to work on.
  </Step>

  <Step title="Open AI agent">
    In the left sidebar, under **Main menu**, click **AI agent**.
  </Step>
</Steps>

<Frame caption="The AI agent screen: where you create and configure your agent(s).">
  <img src="https://mintcdn.com/chatwize/FsRbfI6a4aM0maQD/images/en-agent.png?fit=max&auto=format&n=FsRbfI6a4aM0maQD&q=85&s=81780e3d4e79b82649a66bd7c3fffbb5" alt="Chatwize AI agent page with the option to create an agent" width="2880" height="1800" data-path="images/en-agent.png" />
</Frame>

## What you set

<CardGroup cols={2}>
  <Card title="Instructions (system prompt)" icon="scroll">
    The standing brief for your bot: who it is, what it does and doesn't do, and
    how it handles questions it can't answer.
  </Card>

  <Card title="Tone of voice" icon="comment">
    Formal or informal, brief or expansive.
  </Card>

  <Card title="Model level" icon="microchip">
    How capable the model is and how much context it carries. A higher level
    reasons better about hard questions and costs more credits per message.
  </Card>

  <Card title="Creativity" icon="wand-magic-sparkles">
    Lower = more factual and consistent. Higher = freer and more varied.
  </Card>
</CardGroup>

<Tip>
  Be specific in your instructions about what the bot should do when it doesn't
  know, for example refer to a person or ask for an email address. That stops it
  from guessing.
</Tip>

## External lookups

On this page you also set whether your agent may call an **external service**
during a conversation, for example to fetch an order status. You specify which
addresses it's allowed to reach. The screen shows whether your current model
level supports it.

<Warning>
  Anything your agent may call, any visitor can indirectly cause it to call. So only
  connect addresses that are safe for public use, and have the service itself check
  whether the requested data belongs to this visitor.
</Warning>

## Multiple specialists

Do you have clearly different jobs (sales, support, billing)? Then you can create
**multiple agents** within one chatbot that route to each other. See
[Multiple agents](/en/multi-agent).

## Trying it out

Use **Test chatbot** in the sidebar to check your settings with real questions,
before you go live.