> ## Documentation Index
> Fetch the complete documentation index at: https://docs.callivox.com/llms.txt
> Use this file to discover all available pages before exploring further.
# AI agents overview
> What an AI agent is in Callivox, the kinds you can build, the two reply modes, and how an agent gets attached to a channel.
An AI agent is a configured assistant that reads a conversation and produces a reply. You give it
a name, a set of standing instructions, and a few rules about when it should step aside — Callivox
does the rest, on every channel you connect it to.
## Where agents live
Agents sit in the **Agents** section of the sidebar:
| Page | What it holds |
| ------------------- | -------------------------------------------------------------------- |
| **Chatting Agents** | Customer-facing agents that reply in text on your messaging channels |
| **Voice Agents** | Agents that answer and place phone calls |
| **Assist Agents** | Internal helpers that draft replies for your team inside the inbox |
| **Teams** | Groups of agents that hand conversations to each other |
The whole AI section is **Admin only**, and it appears only when your plan includes AI. **Voice
Agents** needs a plan with voice as well. See [Plans](/billing/plans) for what your plan covers.
## The three kinds of agent
Customer-facing. They reply to customers in text on WhatsApp, Telegram, Instagram, Facebook
Messenger, TikTok, email and the website widget.
Customer-facing. They answer inbound calls or place outbound ones, speaking to the customer.
See [Voice agents](/voice-agents/overview).
Internal. They never message a customer — they summarize the conversation and draft replies
for your human agents. See [Assist agents](/ai-agents/assist-agents).
## The two reply modes (and the hybrid in between)
A chatting agent replies in one of two ways. You pick which when you create it.
"Free-form replies driven by a system prompt." The agent reads the conversation and writes its
own reply, guided by the instructions you give it. Best when questions are open-ended.
"Script the conversation step by step on a canvas." The agent follows a flow you draw — send
a message, ask a question, branch on the answer. Best when the conversation has fixed steps.
A Flow Agent also has a switch on its **General** tab labelled **Let AI handle off-script
messages**. Turning it on gives you the hybrid of the two:
> When on, the agent follows your flow but falls back to the AI for anything the flow doesn't
> cover (Hybrid). When off, it sticks strictly to the flow.
Use hybrid when your process is mostly scripted but customers still ask things you didn't plan
for. See [Flows](/ai-agents/flows).
## Effort, not models
You never choose a provider or a model. You choose an **effort** level — **Low**, **Medium** or
**High** — and Callivox runs that tier on the best available model. Each level has its own
per-conversation credit cost. See [Creating an agent](/ai-agents/creating-an-agent) and
[How credits are spent](/billing/how-credits-are-spent).
## How an agent gets attached to a channel
Creating an agent does not put it to work. An agent replies only where you **bind** it — a binding
says "this agent answers here". Bindings live on the **Bindings** tab of the agent editor.
The tab header shows the current count, for example **Bindings (2)**.
Leave the mode on **Single agent**. (Use **Bind a team** to connect a channel to a whole
[team](/ai-agents/teams) instead.)
**Tenant (everyone)** makes the agent your workspace-wide default. **Channel** binds it to one
connected channel. **Folder** binds it to one inbox folder.
Priority decides which binding wins when more than one could match — the lower the number, the
stronger the binding. New single-agent bindings default to `100`.
A scope can hold only one binding. If a channel or folder is already bound, remove the existing
binding before pointing it at a different agent.
When a message arrives, Callivox picks the answering agent in this order:
1. An agent pinned to that one conversation from the inbox.
2. The default agent of the folder the conversation sits in.
3. The binding on the conversation's channel (lowest priority number wins).
4. The workspace-wide **Tenant** binding.
Bindings scoped to a single conversation appear in the table as **Read-only** — you manage those
from the AI pill inside the conversation. See [AI in the inbox](/inbox/ai-in-the-inbox).
To stop an agent replying everywhere at once without deleting it, turn off **Enabled** on its
**General** tab. A disabled agent shows an amber **Disabled** pill next to its name.
## Next steps
Name it, pick an effort level, and test it in one turn.
Give the agent its instructions, your documents, and the actions it can take.
Decide when the AI should stop and bring in a person.
Put an agent on the phone.