# Agents

> Every door is answered by an agent — an AI brain with its own persona, voice, tools, and guardrails, drawing on the business's shared knowledge. Here's how the roster works.

An **agent** is the AI brain behind a door. When you add a website chat or receptionist service, its backing agent is created automatically; the business keeps a roster of them, and channels pick which agent answers.

## What an agent owns

| Piece | What it controls |
| --- | --- |
| Persona | Name, role description, tone, and the standing instructions — who the AI is when it answers. |
| [Voice](/docs/ai/voice) | For phone agents: the TTS voice, language, and speech behavior. |
| [Knowledge](/docs/ai/knowledge) | What it answers from — the business's trained sources. |
| [Conversation flows](/docs/workflows/overview) | The routed scripts it can follow — one agent carries many, chosen by "use this when…" conditions. |
| [Tools](/docs/ai/tools) | What it can *do* mid-conversation: capture a lead, book a slot, call a connector, open a ticket. |
| [Guardrails](/docs/ai/guardrails) | What it must never do — topics, promises, and behaviors that are off-limits. |

## The roster, drafts, and going live

The business's **Agents** page shows every agent and which service uses it. You can prepare a **draft** agent — new persona, new flows — test it, and swap it live in one move; phone numbers re-wire to the new agent atomically, so there's never a moment when calls hit a half-configured brain.

> **THE MODEL IS MANAGED:** Which LLM powers an agent is administered at the platform level — tenants configure the persona, knowledge, and tools, not the model. You never manage API keys for the AI itself.

## One brain, many doors — or one each

Small businesses usually run one agent across chat and social and a second for the phone (voice needs different pacing). Larger setups split further — a booking agent, a support agent — and route channels to each. Channels are where analytics and logs live; agents are where setup lives.