> For the full page index, fetch https://echonova.ai/llms.txt

# What is an agent

An agent is one saved configuration that answers a caller, holding its persona, its model, its voice, its knowledge, its tools and the channels it is reachable on.

An agent is one saved configuration in your workspace. It holds what the agent
says, which model thinks for it, which voice it speaks with, what it is allowed
to read, what it is allowed to do, and where people can reach it. An agent costs
you nothing until a conversation starts; every conversation it has uses the
settings saved on it at that moment.

## What an agent is made of

Each part is edited on its own tab, so you can change any one of them without
touching the rest.

| Part         | What it decides                                                                              | Where you set it                     |
| ------------ | -------------------------------------------------------------------------------------------- | ------------------------------------ |
| Persona      | The agent's name, the description your colleagues see, its instructions and its opening line | **Persona** tab                      |
| Model        | Which model answers, and how hard it reasons before it replies                               | **Voice** tab, under **Model**       |
| Voice        | The voice it speaks with, per language, and the speech settings for that voice               | **Voice** tab, under **Used Voices** |
| Knowledge    | Which workspace sources it may read, and how much it pulls per turn                          | **Knowledge** tab                    |
| Tools        | What it can do beyond talking                                                                | **Tools** tab                        |
| Availability | The time zone it works in and the hours it answers                                           | **Settings** tab                     |
| Channels     | Where people reach it; see [Channels overview](/docs/agents/channels-overview)               | **Channels** tab                     |

An agent always carries at least one voice, and exactly one voice in that list
is the default. Through the agents API, a list where no entry is flagged
promotes the first one, and a list with two flagged entries is refused with
`Exactly one voice can be the default`.

Knowledge comes in two kinds: an uploaded document, and a web page or crawled
site. Tools come in three kinds, `webhook`, `integration` and `openapi`, plus
the three built-in tools `web_search`, `dialect_detection` and
`end_conversation`.

The channels an agent can be reached on are `whatsapp`, `messenger`,
`instagram`, `slack`, `sms`, `twilio`, `web_voice`, `web_chat`,
`google_calendar`, `zendesk`, `zoho_desk`, `intercom` and `genesys`.
[Channels overview](/docs/agents/channels-overview) says what each one
carries and how it is connected; `web_voice` and `web_chat` are the
playground, and `google_calendar` is the
[Google Calendar](/docs/agents/google-calendar) integration, which gives the
agent calendar tools rather than a place to be reached.

## An agent is either live or archived

There are two states and no third one. The agent detail header shows a live
agent as **Live** and an archived agent as **Archived**, and the button beside
it reads **Pause Agent** or **Restore**. Archiving is reversible, and an
archived agent no longer counts against your plan's active-agent limit.

Archiving stops the sessions you start from the portal. A channel bound to an
archived agent still answers an inbound message or call, and those conversations
are charged. To stop a channel, disconnect it or bind it to another agent.
[Agent status](/docs/agents/agent-status) lists what archiving covers, and
[Channels overview](/docs/agents/channels-overview) how a channel is bound.

## Sessions and turns

A **session** is one conversation with one agent. It is a `voice` or a `text`
conversation, and its status is `active` while it runs, then `ended`, `failed`
or `out_of_credits`. Each conversation records the agent, the channel, when it
started, when it ended, and who the contact was.

A **turn** is one exchange inside a conversation: the caller says something, the
agent answers. Turns are what you are charged for. Each conversation keeps a
running count of its turns, the model's prompt and completion tokens, the
characters spoken, the seconds of audio transcribed, the messages in and out,
the failed sends, the response time and the tool calls.

## What it costs

An idle agent costs nothing. A call is charged 2,500 credits for every minute it is connected, and every turn is charged for its AI usage as the conversation runs.
[What a turn costs](/docs/agents/what-a-turn-costs) breaks down what a conversation is charged for.

## Get it right

One agent per job beats one agent for everything. The prompt, the knowledge and
the tools are all set once per agent, so a single agent covering unrelated jobs
carries instructions that are wrong most of the time.

**Recommended**

```text
Support agent: orders, deliveries, returns. Knowledge: returns policy, delivery FAQ.
Booking agent: appointments only. Tools: calendar booking.
```

**Less effective**

```text
One agent: support, bookings, sales, recruitment. Knowledge: every document in the workspace.
```

## FAQ

### Do I pay for an agent that is sitting there?

No. Credits are charged for conversations: their connected time on a call and their AI usage. An agent with no
conversations costs nothing, though it does count against your plan's
active-agent limit until you archive it.

### Can two agents share the same voice or the same document?

Yes. Voices and knowledge sources belong to the workspace, and any number of
agents can use them. Each agent keeps its own tuning and its own selection.

### Does changing a setting affect a call that is running?

No. A conversation finishes on the settings it started with. Start a new
conversation to hear a change.

### How many agents can I have?

As many as your plan allows live at once. Archive the ones you are not using;
archived agents do not count.

### Can one agent answer on several channels?

Yes. An agent can be bound to several channels, and every one of them gets the
same persona, model, voice, knowledge and tools. See
[Channels overview](/docs/agents/channels-overview).

## Related

Start here for each part of an agent:

- [Create an agent](/docs/agents/create-an-agent): make your first one in a few minutes
- [Write the persona](/docs/agents/persona): the text that decides how it speaks
- [How knowledge works](/docs/agents/knowledge-overview): what it is allowed to read
- [Tools overview](/docs/agents/tools-overview): what it can do beyond talking
- [Channels overview](/docs/agents/channels-overview): where people reach it
- [Find and read a conversation](/docs/agents/conversations): see what it actually said
- [Data retention](/docs/agents/data-retention): what is kept and for how long

And for the rest of this page:

- [Build an agent with AI](/docs/agents/build-with-ai): answer six questions instead of filling in tabs
- [The agent overview tab](/docs/agents/agent-overview-page): the numbers an agent opens on
- [Agent status](/docs/agents/agent-status): pausing, restoring and what each one stops
- [What a turn costs](/docs/agents/what-a-turn-costs): how a conversation is priced
- [What is EchoNova](/docs/get-started/what-is-echonova): the rest of the product