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

# Agent Workforce

> Workers that communicate and act, spun up per task - one agent, one step, one entity.

The Agent Workforce is workers that communicate and act. Each worker is an
agent - voice, action, chat or browser - that runs against the instructions
it was given, on one step, for one entity, spun up per task and torn down
when the task completes. Not every kind ships today - see
[Status](#status). There is no long-lived agent process holding state
between customers; state lives on the spine, workers are disposable.

**Tasks are the unit of work.** A task says: run this agent, with this
instruction set, for this entity, now. Calling already runs this way - every
outbound call is a task dispatched to a voice agent - through the
[Task API](https://docs.unpod.ai/api/execution/task-make-call-overview.md).
Human handoff is built in: an agent that hits an escalation condition
transfers the call to a person instead of improvising. Guardrails today are
playbook-level - `never_say` lists and hard gates - and a policy layer is in
build.

You build these workers with [SuperDialog](https://docs.unpod.ai/superdialog/introduction.md), the
open-source agent framework, and a
[playbook](https://docs.unpod.ai/agent-studio/what-is-a-playbook.md) is the instruction set a worker
runs against. SuperDialog's defining rule -
[the wire carries text, not audio](https://docs.unpod.ai/superdialog/introduction.md) - is what lets
the same playbook drive a phone call, a chat thread, or a browser session.

## The primitives

| Primitive | What it is | Where it lives today |
| - | - | - |
| Agents | A worker built on SuperDialog, spun up per task | [Build an agent](https://docs.unpod.ai/agent-studio/build-an-agent.md) |
| Tools | Actions a worker can take mid-conversation | [SuperDialog tools](https://docs.unpod.ai/superdialog/tools.md) |
| Policies | Guardrails on what a worker may say and do | `never_say` and gates in [playbooks](https://docs.unpod.ai/superdialog/playbooks.md); policy layer in build |
| Handoffs | Escalation to a human, or to another agent, mid-interaction | [Call lifecycle](https://docs.unpod.ai/agent-sdk/call-lifecycle.md) (voice) |

## On the spine

A worker reads context and the entity it was dispatched for, then acts:
each task produces an interaction on the entity's conversation, and each
interaction closes with a structured outcome. Task dispatch, tool calls,
handoffs and completion all emit events. Workers never write to the spine
directly outside their task - what a worker learned reaches the Context
Engine through the outcome, not through a side channel.

## What stays hidden

Model-provider orchestration internals stay hidden: which LLM served a
given turn, provider failover, prompt assembly, and inference routing.
You declare the playbook and the tools; the workforce decides how to run
them. A provider outage is Unpod's problem, not a change to your agent.

## Status

Production. Voice agents carry production traffic today, spun up per task,
with human handoff. Actions run today as [tools](https://docs.unpod.ai/superdialog/tools.md) inside
voice agents. Browser agents are in build. Chat runs through the same
workers via the [Chat API](https://docs.unpod.ai/get-started/chat).

## Go deeper

> [card] [SuperDialog](https://docs.unpod.ai/superdialog/introduction.md)
>
> The open-source framework you build workers with.

> [card] [What is a playbook](https://docs.unpod.ai/agent-studio/what-is-a-playbook.md)
>
> The instruction set a worker runs against.

> [card] [Build an agent](https://docs.unpod.ai/agent-studio/build-an-agent.md)
>
> From a blank playbook to a running worker.

> [card] [Task API](https://docs.unpod.ai/api/execution/task-make-call-overview.md)
>
> Dispatch a task - calling runs this way today.

> [card] [Flows & Journeys](https://docs.unpod.ai/agent-studio/flows-journeys.md)
>
> The workflows that dispatch tasks to the workforce.


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