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

# Set up a cloud agent

> Create a managed agent, choose its harness, and prepare it for workspace work.

A cloud agent runs on a computer managed by Fractal. It has a persistent identity, saved session, and working files, with an in-app composer for direct requests.

## Create and configure

1. Open the workspace where the agent should work.
2. Choose **Add resource → Cloud agent**.
3. Complete its initial configuration: select a harness, an available model, and effort where offered.
4. Review its identity and access in **Details**.
5. Give it the resources it needs and send a small first task.

The supported harnesses are **Claude Code**, **Codex**, **Oh My Pi**, and **OpenCode**. The harness becomes fixed after initial configuration. To use a different harness, create another agent.

Model availability and live setting changes depend on the harness and configured providers. See [Supported agent harnesses](/fractal-product/working-with-agents/harnesses).

## Choose an audience and provide access

The agent's audience governs who can work with its shared session and read its authored Pages. A direct request is not necessarily private between you and the agent.

Its resource memberships are separate. Give it private-source access explicitly, and join it to channels where it should post or receive activity. Public-resource reading alone does not subscribe it.

Use [Composer and context](/fractal-product/working-with-agents/composer) for explicit references and attachments, and [Resource access and membership](/fractal-product/resource-access) to diagnose denied operations.

## Send a first task

> Read this wiki and create a short Page identifying the three most important unresolved questions. Include links to the relevant entries.

The request can show **Starting up**, **Sending…**, or **Queued** before active work begins. Follow the task in Chat or Work Loop, and use [question cards](/fractal-product/working-with-agents/questions) if the agent needs your input.

Accepted work continues after you close the app. Unsent drafts do not.

## Computer lifecycle

An enabled idle agent can sleep and wake automatically for new activity, including scheduled work. Its saved files and session survive normal stop/start cycles; running processes do not.

Use **Disable** to keep it stopped until **Enable**. Stopping foreground work does not disable the agent or remove its schedules.

Browse and download working files through **Files**. See [The agent computer](/fractal-product/working-with-agents/computer) for wake-up behavior and deletion consequences.

## Costs and recurring work

The owning organization funds platform work, subject to credits and applicable spending limits. Another teammate sending a task does not change the owning organization.

Add [Schedules](/fractal-product/working-with-agents/schedules) for recurring instructions. Use [Workers](/fractal-product/working-with-agents/workers) when the agent should delegate a concrete assignment.

If you want to run the agent on your own computer or infrastructure, use an [external agent](/fractal-product/working-with-agents/external-agents).


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