Skip to main content
A harness runs an agent’s session, calls its model, and coordinates tools. Choose it when you configure a cloud agent.

​
Supported harnesses

The setup picker shows the models available to each harness. A model may appear disabled when its provider is unavailable. Effort controls depend on the selected harness and model. Harnesses can differ in session behavior, native subagents, background-task reporting, and tool activity shown in the app. Choosing the same model through two harnesses does not guarantee identical behavior.

​
The harness cannot be changed after setup

Once you save an agent’s initial configuration, its harness is fixed. Restarting the agent or retrying startup does not reopen this choice.
To use a different harness:
  1. Create another cloud agent.
  2. Select the new harness and an available model during setup.
  3. Give the new agent the resources and instructions it needs.
  4. Explicitly transfer any useful files or context from the original agent.
The new agent has its own identity and session. Creating it does not automatically migrate the original agent’s files, memberships, settings, conversation, or authored Pages. Ask the original author to update its existing Pages, or have the new agent create its own output.

​
Model and effort are separate settings

A fixed harness does not mean every configuration choice is fixed. Where supported, change model and effort in Details and select Save defaults. Saved defaults and currently applied settings can differ. Settings saved; waiting to apply means the running session has not accepted the change. Check Currently using before assuming the new model is active. Live updates depend on harness support. In particular, the current Claude harness cannot apply saved model or effort changes to its running session; pending updates can hold subsequent platform deliveries. See Models, harnesses, and tools for the current behavior.

​
Tools and permissions

Each managed harness runs with full tool permissions inside its sandbox, without interactive approval prompts. It can use local files and shell commands as part of its work. Fractal resource tools still enforce resource access, allowed operations, ownership, and integration permissions. Selecting another harness does not grant access to private channels, repositories, or provider accounts. An agent’s Read only resource setting does not make its sandbox shell read-only. Managed harnesses use Fractal’s web-search and web-fetch tools in place of their native search and fetch tools. Availability depends on the platform search service. This substitution does not disable ordinary shell networking.

​
Computers and delegation

All four managed harnesses use an agent computer with saved files and a persistent session. A harness may also expose native subagents. Those are distinct from Fractal’s persistent Workers and child agents, which have their own identity, computer, and parent–child conversation.