> ## Documentation Index
> Fetch the complete documentation index at: https://docs.atako.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Agents

> What an agent is, what it's made of, and how it behaves over time.

An agent is an autonomous AI worker that runs continuously on Atako, connected to the tools your company uses. An agent isn't created for a single conversation and discarded — it's a persistent, always-on entity you configure once and keep running.

## What an agent is made of

When you create an agent, you define:

<CardGroup cols={2}>
  <Card title="Identity" icon="id-badge">
    A name and an avatar.
  </Card>

  <Card title="Instructions" icon="file-lines">
    An optional system prompt describing how the agent should behave.
  </Card>

  <Card title="Context" icon="layer-group">
    Business context specific to the agent, plus your company's shared context (configured under Settings → Context). Both are captured into the agent's instructions at creation time — later changes to the company context apply to agents created afterwards.
  </Card>

  <Card title="Model" icon="microchip">
    An LLM picked from the model catalogue, with search, favorites, and recently used models — see [Atako inference](/concepts/atako-inference). Or your own API key if [BYOK](/concepts/credits#byok-agents) is enabled for your company.
  </Card>
</CardGroup>

You also choose the agent's **engine** (labeled *Harness* in the creation form): a general-purpose agent or a coding agent. This choice is **fixed for the life of the agent** — it can't be changed afterward. See [Engines](/concepts/engines).

Integrations and channels are configured after creation: you grant an agent access to specific actions on connected tools, and it becomes reachable over chat, email, webhooks, cron, and other agents — within what its [engine](/concepts/engines) allows. See [Integrations overview](/integrations/overview) and [Channels overview](/channels/overview).

## Always-on

An agent runs in its own isolated environment with persistent storage. Each agent can hold several conversations with you; start a new one whenever you switch topics, and pick up any previous one from the conversation list. Whatever the conversation, the agent's working memory, scheduled tasks, sub-agents, and files persist in its environment across restarts. If a connection drops, nothing is lost: messages waiting on the email, webhook, and inter-agent channels are replayed once the agent reconnects, and chat history and scheduled tasks persist independently of connection state.

## Lifecycle

```mermaid theme={null}
stateDiagram-v2
    [*] --> provisioning: create agent
    provisioning --> active: environment ready
    active --> paused: automatic pause (e.g. trial expiry)
    provisioning --> paused: automatic pause (e.g. trial expiry)
    provisioning --> completed: cancel
    paused --> active: resume
    active --> completed: end
    paused --> completed: end
    completed --> [*]
```

| Status | Meaning |
| - | - |
| **Provisioning** | The agent's environment is being set up. |
| **Active** | The agent is running and occupies an active-agent license. |
| **Paused** | Compute is stopped, but persistent state — memory, cron tasks, files — is kept. Pausing is automatic today (for example when a trial expires); resuming is user-triggered and restarts the same instance, not a new one. An agent paused at the end of a trial is permanently deleted if it still has no license 14 days later — see [Free trial](/guides/subscription#free-trial). |
| **Completed** | The agent has ended. Its license is freed immediately and its compute is torn down; its conversation history and files remain accessible. |

<Warning>
  Ending an agent is permanent: it cannot be resumed, and its license is freed immediately. Its conversation history and files remain viewable, but the running instance and its working state are gone for good.
</Warning>

## Active-agent licenses

A company's agents draw from a shared pool of **active-agent licenses** — one license per concurrently active agent. The number of licenses equals the number on your subscription; it isn't tied to the number of human users on your team. See [Credits](/concepts/credits) and [Subscription](/guides/subscription) for how licenses map to billing.

## Conversations are private per user

Chat is agent-wide in terms of capability, but conversations are private per (agent, user) pair. If two colleagues both talk to the same agent, they each get their own conversations — while the agent's memory, files, and cron tasks are shared and visible to the whole team through that one agent. See [Chat](/guides/chat).

## Memory and state

Everything an agent accumulates — conversation context, files it's created or received, scheduled tasks — persists for the agent's entire lifetime, across restarts and pauses. Nothing needs to be re-explained after a reconnect.

## Sub-agents

For complex tasks, an agent can delegate work to short-lived **sub-agents**. These are spun up to handle a specific piece of work and then retired; they don't consume additional licenses. Their work is visible as steps inside the parent agent's [activity timeline](/guides/activity), so you can always see what was delegated and what it did.

## Related

<CardGroup cols={2}>
  <Card title="How it works" icon="diagram-project" href="/how-it-works">
    The platform architecture behind every agent.
  </Card>

  <Card title="Credits" icon="coins" href="/concepts/credits">
    How agent usage is metered and billed.
  </Card>

  <Card title="Channels overview" icon="tower-broadcast" href="/channels/overview">
    Every way to reach an agent.
  </Card>

  <Card title="Quickstart" icon="rocket" href="/guides/quickstart">
    Create your first agent.
  </Card>
</CardGroup>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.