Creating an agent

You build and configure customer agents under Conversations → Manage agents. Each agent is a self-contained configuration: a persona, the conventions it follows, the knowledge it can read, the actions it can take, and the channels it answers on.

​

The workspace

Open Manage agents and create or select an agent. Its configuration is organized into tabs:

​

Persona

The persona is the agent’s system prompt—its identity and behavior. Write it the way you’d brief a new teammate: who they are, who they’re talking to, what they do, and what matters. Keep the persona about identity and behavior, not phrasing. Specific, repeatable rules (“always greet by first name”, “never quote prices on LINE”) belong in guidance cards, which keep the persona short and let you scope each rule to the right channel. How the agent sounds—tone, register, sign-off—belongs in Voice & tone, below.

​

Voice & tone

By default the agent writes in its own voice, guided by the persona and any Voice & tone guidance cards. For one consistent brand voice across every reply, turn on Voice & tone rewrite to separate what the agent says from how it sounds:

• Describe your brand voice in its own field—tone, register, emoji policy, sign-off, do’s and don’ts.
• The agent then writes plain, correct substance, and a separate fast pass rewords every reply into that voice just before it’s staged or sent.

The rewrite changes wording only. Facts, numbers, prices, names, links, and bubble splits (---) are preserved exactly—and if the pass can’t guarantee that, the original reply goes out unchanged. The reworded text is what your team sees in a draft and what the customer receives.

​

Knowledge

An agent only knows what you grant it:

• Knowledge store — the agent reads from your org’s knowledge store within the bounds of its assigned role. The role determines which folders it can see, so two agents can draw on different knowledge.
• Data stores — grant individual data stores the agent may query. Customer-facing reads are further limited to the identity each conversation is scoped to.
• Web search (toggled under Behavior) — enable it when the agent needs current, public information it can’t get from your own knowledge. Off by default; keep it off for agents that should answer only from your own knowledge.

​

Actions

Actions let an agent do things, not just answer—look up an order, create a ticket, check availability. An action is an endpoint on one of your apps that the agent is allowed to call. You grant actions two ways:

• Specific endpoints — pick individual endpoints the agent may call.
• Whole apps — grant an app, and the agent can call any of its agent-callable endpoints (including ones you add later).

Only endpoints that opt in to agent calls are available, and access is re-checked on every run—so revoking it on the app side takes effect immediately. See Cross-App Communication for how endpoints opt in.

​

Settings

• Reasoning effort — how much the model deliberates before answering. Higher is slower but more thorough; leave it on the profile default unless you have a reason to change it.
• Status — switch an agent to disabled to stop it replying everywhere (including test chats) without deleting it.

Every agent runs on the same managed platform model—there’s no per-agent model to pick.

​

Reply mode

Set the agent’s default reply mode—auto-send, auto-draft, or off—under Behavior. This is the default for new conversations; your team can override it per conversation at any time.

​

Reply style

Also under Behavior, choose how replies are shaped:

• Reply style — short chat bubbles (the default) splits a reply into separate messages wherever the agent puts --- on its own line, which reads naturally on LINE and chat; single message always sends one block.
• Reply delay — how long to wait after a customer message before replying, so a rapid burst of messages batches into one response. Blank uses the org default; 0 replies immediately.

​

Pair it to a channel

A configured agent doesn’t answer anything until a channel routes to it. Open Conversations → Channels and pair each connected channel to an agent:

• LINE and Facebook accounts pair here—inbound on that connection routes to the agent you choose. (Instagram DMs come through the Facebook Page connection.)
• Web needs no pairing—your app names the agent when it calls the SDK.

Connect the LINE or Facebook accounts themselves first under Settings → Integrations. See LINE and the Connected Services overview for connecting accounts.

​

Test before you go live

Use the Test tab to hold a sandbox conversation with the agent—same persona, guidance, knowledge, and actions, but nothing reaches a real customer. Test conversations are kept separate from the live pool.

​

Next steps