| name | subagent-launcher |
| description | Start or delegate work to a configured project subagent. Use when a task matches an available subagent, or when a platform-specific subagent command is needed. |
Subagent Launcher
Workflow
- Select: Identify the best subagent for the current task.
- Resolve Config: Use standard resolution priority (check workspace
agents/, then global config, then native roles). Use native role metadata if available rather than inventing a file path.
- Load Skills: Resolve every skill in the config's
skills field via the platform registry. If supported, attach them natively. Otherwise, instruct the subagent to read the resolved skill paths before acting. Report any missing skills.
- Delegate: Pass the exact task, context, and explicitly provided constraints. Do not do the work yourself if the subagent owns it.
Antigravity (define_subagent / invoke_subagent)
- Define the subagent using the resolved config (or native role). Grant full permissions (
enable_write_tools, enable_mcp_tools, enable_subagent_tools) unless restricted by config.
- Build the system prompt from the config's
instructions and attach the resolved skills by path.
- Invoke with the user's task.
Codex (multi_agent_v1.spawn_agent)
- Prefer spawning by native role name. Use generic
worker/explorer only if no custom role exists.
- Set
fork_context: false unless history is needed.
- Pass the config identifier and skills as structured items:
{
"agent_type": "<role-or-worker>",
"fork_context": false,
"items": [
{ "type": "text", "text": "Use config for subagent <name>" },
{ "type": "skill", "name": "<skill-name>" },
{ "type": "text", "text": "Task: <USER_REQUEST>" }
]
}
- If native items are unavailable, write a prompt listing required skills/paths for the worker to follow.
- Close completed agents when no longer needed.
Reporting
Report the selected subagent, execution status, and final result. Do not claim success until it finishes.