Guides
Define and install subagents
A subagent is a role definition with a usage description, role prompt, and Skill list. Creating one does not install it. Enable its required Skills first, then let each Agent load the role through its native mechanism.
Last reviewed
Define a role
When to use this
Use a subagent when recurring work needs its own role instructions and Skill dependencies.
A subagent is a role definition with a name, usage description, role prompt, and Skill references. The description explains when to use it; the prompt defines its behavior. The prompt and Skill list can be empty to use native defaults.
- Open Subagents and choose Create subagent.
- Enter a name and usage description. Optionally add a role prompt and select Skills from the central library.
- Save the definition. Creating it does not install it or change Skill switches.
Prepare dependencies and install
Enable the selected Skills for the target Agent in the Skills page and check that their entries are healthy. Open the subagent detail and choose Install for that Agent. One role can be installed in several Agents, and an Agent can have several roles. Codex dependencies continue to use .codex/skills.
Claude Code preloads selected Skills; OpenCode uses native Skill permissions; Codex uses native configuration layers. OpenClaw uses a separate Agent and workspace. Prepare its native Agent roster and default ownership, then grant delegation permission to the caller. The product does not change routing or widen permissions.
Installed configuration confirms that files were written. Follow the reload or new-session guidance and verify the role with the native tool. Each Agent retains its own loading and invocation behavior.
Update and uninstall
Saving an installed definition updates all managed installations. Review the affected Agents first. Unknown occupants or external edits block the complete update.
Uninstalling removes the role's managed configuration and preserves Skill switches. Deleting a role uninstalls it everywhere; OpenClaw sessions, memory, and runtime files are retained.
You can disable a Skill directly; the subagent detail then reports a dependency issue. A central Skill with subagent references cannot be removed. Edit the roles to remove those references first.