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

# Getting started

> Create, test, and activate your first AI Agent.

Use this guide to turn one customer-support job into a tested AI Agent.

## Before you begin

Connect the channel and business systems the agent needs. Then decide:

* which customer problem the agent should handle
* which channels it should work on
* what should start it
* which data and actions it needs
* when it should ask for human approval
* when it should escalate

## Create an AI Agent

Ask Torben turns your support goal into a draft with instructions, triggers, skills, actions, and tests.

<Steps>
  <Step title="Open AI Agents">
    Go to **AI** → **AI Agents**, or [open the AI Agents overview](https://armin.cx/app/_/ai/agents-v2).
  </Step>

  <Step title="Start with Ask Torben">
    Click **New agent**, choose a template, or [open Ask Torben](https://armin.cx/app/_/ai/agents-v2?askTorben=open). Edit the prompt before sending if needed.
  </Step>

  <Step title="Describe the job">
    Explain what customers ask, what a successful outcome looks like, and when a person must take over.
  </Step>

  <Step title="Review the draft">
    Check the instructions, triggers, skills, actions, and approval settings. Connect any missing integration or [MCP connection](https://armin.cx/app/_/ai/agents-v2/mcp).
  </Step>

  <Step title="Choose a model">
    Use **GPT 5.6 Luna**. It is the recommended default and is sufficient for supported customer-support use cases, including multi-step jobs with lookups, updates, delays, follow-ups, and escalation.
  </Step>

  <Step title="Test with real scenarios">
    Test realistic happy paths, missing data, exceptions, and messages that should go to a person.
  </Step>

  <Step title="Activate gradually">
    Require approval for sensitive actions at first. Review [run history](https://armin.cx/app/_/ai/agents-v2/runs) before giving the agent more autonomy.
  </Step>
</Steps>

Agent-specific links contain private route IDs. Open the agent from the [overview](https://armin.cx/app/_/ai/agents-v2), then use its **Runs** and **Tests** tabs.

## Best practices

* Start with one clear job per agent.
* Combine exact filters with a clear description of customer intent.
* Give the agent explicit escalation rules.
* Require approval for refunds, cancellations, and other sensitive changes.
* Use tests for both expected and edge-case customer messages.
* Review early runs before enabling more automatic behavior.

<Tip>
  Ask Torben can [diagnose a run and improve the agent](https://armin.cx/app/_/ai/agents-v2?askTorben=Help%20me%20diagnose%20an%20AI%20Agent%20run%20and%20suggest%20specific%20improvements.%20Ask%20me%20for%20the%20agent%2C%20ticket%2C%20or%20run.) when a test or live ticket behaves unexpectedly.
</Tip>

Next, learn how [routing and conversational ownership](/ai-agents/routing-and-ownership) work.