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

# Starcat Customer Agents

> Put a Starcat agent on your LINE, Facebook, Instagram, or web channels to answer customers automatically—with your team reviewing or taking over whenever they want.

<Info>
  **Starcat Customer Agents is in early access.** Look for **Conversations** in your dashboard
  sidebar. If you don't see it, [contact support](mailto:support@stardeck.ai) to request access.
</Info>

Where [Starcat](/starcat) works *on your org*—building apps, writing knowledge, running tasks—a **customer agent** works *with your customers*. It answers them on the channels you connect (LINE, Facebook, Instagram, or your own app over the web SDK): it reads each incoming message, decides whether to reply, and either sends a reply or stages one for your team to approve.

Every conversation lands in one shared **Conversation Pool**, so your team sees every thread in one place no matter which agent or channel it came from—and can step in, edit a draft, or take the conversation over at any point.

```mermaid theme={null}
flowchart LR
    Cust["💬 Customer"] -- "message" --> SD["☁️ Stardeck"]
    SD -- "routes to the paired agent" --> A["🐈 Starcat customer agent"]
    A -- "reply / draft / handoff" --> Pool["📥 Conversation Pool"]
    Pool -- "auto-send or operator approves" --> SD
    SD -- "delivers on the channel" --> Cust
    Op["🧑 Your team"] -- "review · edit · take over" --> Pool
```

## Where it lives

Open **Conversations** in your dashboard sidebar. From there:

* **Conversations** — the shared pool of every thread across all channels and agents. This is your team's inbox.
* **Channels** — pair each connected channel (a LINE account, a Facebook Page) to the agent that should answer it.
* **Manage agents** — create and configure customer agents: persona, behavior, knowledge, actions, and settings.

## How an agent answers

When a customer message arrives on a paired channel:

1. **Classify** — the agent first decides whether this message even warrants a reply (a "thanks!" usually doesn't). See [the reply classifier](/customer-agents/operator-workflow#the-reply-classifier).
2. **Compose** — if it should reply, it drafts an answer shaped by its persona, its [guidance cards](/customer-agents/guidance-cards), and any [playbook](/customer-agents/playbooks) you've set—drawing on the knowledge and actions you've granted it.
3. **Deliver** — depending on the agent's reply mode, the answer is either sent straight to the customer or staged as a draft for your team to approve.

## Reply modes

Each agent has a default **reply mode**, and you can override it on any single conversation without changing the default.

| Mode           | What happens                                                                          | Use it when                                         |
| -------------- | ------------------------------------------------------------------------------------- | --------------------------------------------------- |
| **Auto-send**  | Replies go straight to the customer with no review.                                   | You trust the agent on this channel.                |
| **Auto-draft** | Replies are staged as pending drafts; an operator approves or edits each before send. | You want a human in the loop on every reply.        |
| **Off**        | The agent runs but never replies.                                                     | You're testing, or you've paused automated replies. |

<Note>
  Reply mode controls **delivery**, not whether the agent thinks. Even in **Off** mode the
  conversation still appears in the pool—your team just handles it manually.
</Note>

## Channels

A customer agent can answer on:

* **LINE** — a connected LINE Official Account (direct messages, groups, rooms)
* **Facebook & Instagram** — Messenger and Instagram DMs on a connected Facebook Page
* **Web** — your own app, by calling the agent directly through the customer SDK

LINE and Facebook are **paired** to an agent under **Channels**: inbound messages on that connection route to the agent you chose. Web conversations don't need a pairing—your app names the agent when it sends. Connect LINE and Facebook accounts first under [Settings → Integrations](https://www.stardeck.ai/dashboard/settings?tab=integrations).

## Your team stays in control

Nothing about a customer agent locks your team out:

* **Review drafts** — in auto-draft mode, approve, edit, or reject each reply.
* **Take over** — switch a single conversation to **Off** and answer it yourself.
* **Receive handoffs** — an agent can [hand a conversation to a human](/customer-agents/operator-workflow#handing-off-to-a-human) on its own—when a playbook reaches its goal, or when the customer asks for a person. Once handed off, the agent stops replying until your team resumes it.

## Next steps

<CardGroup cols={2}>
  <Card title="Create an agent" icon="robot" href="/customer-agents/creating-an-agent">
    Persona, model, channels, and reply mode
  </Card>

  <Card title="Guidance cards" icon="list-check" href="/customer-agents/guidance-cards">
    Reusable conventions that shape every reply
  </Card>

  <Card title="Playbooks" icon="route" href="/customer-agents/playbooks">
    Goal-directed flows that route to a human
  </Card>

  <Card title="Operator workflow" icon="inbox" href="/customer-agents/operator-workflow">
    The pool, draft review, and handoffs
  </Card>
</CardGroup>