> ## 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.
# Operator workflow
> The shared Conversation Pool, reviewing and approving drafts, and taking a conversation over from a customer agent.
Everything a customer agent does is visible to your team and reversible. This page covers the operator side: where conversations live, how draft review works, why an agent sometimes stays quiet, and how handoffs work.
## The Conversation Pool
Open **Conversations** in your dashboard. Every thread—across every channel and every agent—lands in one shared pool. It's your team's inbox.
From the pool you can:
* **Filter and search** by agent, channel, customer, or text.
* **Open a conversation** to see the full thread alongside a side panel of details and actions.
* **Reassign** a conversation to a different agent.
* **Merge** two threads that are really the same customer into one.
* **Bulk-act** on several conversations at once—change their reply mode or assignment together.
A conversation shows when an agent is actively working on a reply, so your team can see it's mid-response before they jump in.
## Reviewing drafts
When an agent runs in [auto-draft mode](/customer-agents/overview#reply-modes), its reply doesn't go out—it's staged as a **pending draft** on the conversation. Each draft gives you:
* **Approve & send** — deliver the reply as-is.
* **Edit, then send** — adjust the wording first; your edit is what the customer receives.
* **Reject** — discard it; the agent doesn't reply.
Auto-send skips this step—replies go straight to the customer. You can move any single conversation to auto-draft (or **off**) without changing the agent's default, when you want a closer eye on a particular thread.
## The reply classifier
An agent doesn't reply to every message. Before composing, a **classifier** decides whether this message warrants a reply at all—so the agent stays quiet on a bare "thanks!" or "ok 👍" instead of replying for the sake of it.
* The decision is recorded **in the thread** as a system line—you can always see whether the agent replied or stayed quiet, and why.
* You shape it two ways: the agent's classifier instruction under **Behavior**, and [scoped rules](/customer-agents/guidance-cards#two-kinds-of-card) (classifier-surface guidance cards) like "always reply when someone mentions a refund."
* An explicit **@mention** of the agent always gets a reply—it bypasses the classifier.
Turn the classifier off (under **Behavior**) and the agent replies to every inbound message.
## Handing off to a human
An agent can route a conversation to a person on its own—when a [playbook](/customer-agents/playbooks) reaches its goal, or when a customer asks for someone. When it does:
* The conversation is **marked as handed off** and the agent **stops replying** on every channel.
* It surfaces in the pool with the route the customer chose (a call, a meeting, your sales team), ready for a person to pick up.
* If the agent had also staged a reply in the same turn, that draft is dropped—handing off means the agent steps back rather than getting the last word.
### Resuming
A handed-off conversation stays with your team until you resume it. Once you've helped the customer—or decided the agent should take it again—clear the handoff to let the agent resume answering. Until then, your replies go out as a normal operator, and the agent stays out of the way.
A handoff is independent of reply mode. It's a hard stop on the agent for that conversation, not a
mode change—so an auto-send agent still goes quiet the moment it hands off.
## Taking over manually
You don't have to wait for a handoff. At any time you can:
* Switch a conversation to **off** and answer it yourself.
* Edit or reject a pending draft.
* Reassign it to a different agent.
The agent never locks your team out of a conversation.
## Next steps
Auto-send, auto-draft, and off
Where agent-initiated handoffs come from
Shape replies and the classifier