> ## 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.

# Skills and actions

> Give AI Agents the knowledge, systems, and permissions they need.

Skills and actions turn an AI Agent from a reply assistant into a support operator that can complete work across your business systems.

## Core concepts

<CardGroup cols={2}>
  <Card title="Instructions" icon="file-text">
    Define the support job, boundaries, tone, and escalation rules.
  </Card>

  <Card title="Skills" icon="book-open">
    Give the agent reusable policies and procedures, such as return eligibility.
  </Card>

  <Card title="Actions" icon="wrench">
    Let the agent retrieve data, reply, update a system, wait, report, or escalate.
  </Card>

  <Card title="Integrations" icon="plug">
    Connect your shop, ERP, fulfillment, helpdesk, and other business systems.
  </Card>

  <Card title="MCP connections" icon="network-wired">
    Extend agents to more than 1,200 available connectors for the systems your support operation
    uses.
  </Card>

  <Card title="Approval" icon="user-check">
    Keep a person in control of sensitive replies and changes.
  </Card>
</CardGroup>

## Connect business systems

Integrations and MCP connections are two ways to give an agent access to live business data and approved operations. Connect the system first, then enable only the actions that the support job requires.

<CardGroup cols={2}>
  <Card title="Integrations guide" icon="plug" href="/integrations/intro">
    Connect supported commerce and support systems.
  </Card>

  <Card title="MCP connections" icon="network-wired" href="https://armin.cx/app/_/ai/agents-v2/mcp">
    Find and connect additional business tools.
  </Card>
</CardGroup>

Common customer-support actions include:

* looking up an order, shipment, return, invoice, or customer
* updating a shipping address or ticket
* creating a refund or discount
* sending a reply, email, or CSAT link
* waiting before a follow-up
* routing to a team or person
* resolving a completed ticket

## Build a focused agent

<Steps>
  <Step title="Write the job in instructions">
    State what the agent owns, how it should speak, when it must stop, and which cases are out of
    scope.
  </Step>

  <Step title="Add reusable support knowledge">
    Add skills for policies or procedures that apply across cases, such as return windows or
    escalation criteria.
  </Step>

  <Step title="Connect the required systems">
    Use an integration or MCP connection for the live data and operations the job needs.
  </Step>

  <Step title="Enable only required actions">
    Keep the action list narrow. Read-only lookups and sensitive changes should be visibly distinct.
  </Step>

  <Step title="Set approvals and test">
    Require approval for high-impact writes, then test success, failure, and escalation cases.
  </Step>
</Steps>

## Hand-off actions

Hand-off actions let an agent give the conversation to the right next owner.

* **Another AI Agent:** The current agent can check which active specialists are eligible for the ticket. It may hand the conversation to one suitable specialist. The specialist takes over, and the first agent stops.
* **A human:** When no specialist fits or judgment is needed, the agent can hand the ticket to a configured person, team, or human queue. This ends the current agent's responsibility.

An agent only uses the hand-off options you configure and describe in its instructions. It will not pass a ticket around endlessly. Repeated or circular hand-offs stop. Cases that still need attention go to humans. Cases with no needed reply or action can end quietly.

A hand-off to a human does not block every AI Agent forever. A later customer message can be matched again. To keep the ticket human-only, narrow or disable matching AI Agent triggers, stop Classic AI auto-reply for the channel, and narrow or disable macro Workflows with AI prompts.

Some cases need no hand-off. A duplicate, an already solved request, or a simple thank-you may end without a reply or human alert when no action is needed.

Read [Routing and ownership](/ai-agents/routing-and-ownership) for the full follow-up and hand-off behavior.

## Approval

Start with approval for:

* refunds
* cancellations
* address updates
* discounts or replacement shipments
* other writes to business systems
* sensitive customer replies

<Warning>
  Do not give an agent broad write access. A returns agent should not receive unrelated account,
  marketing, or catalog actions.
</Warning>

## Best practices

* Keep one agent focused on one support outcome.
* Use current integration data instead of asking customers for information you already have.
* Add approval before activation and loosen it only after reviewing runs.
* Explain failures and route the ticket instead of leaving the customer at a dead end.

Action availability does not decide which agent owns the conversation. Configure selection separately in [Triggers and filters](/ai-agents/triggers-and-filters) and [Routing and ownership](/ai-agents/routing-and-ownership).