| name | subagent-launcher |
| description | Start or delegate work to a configured project subagent and ensure its configured skills are loaded. Use when AGENTS.md or the user says to call, launch, spawn, invoke, delegate to, or use a specialized subagent; when a task matches an available subagent listed in AGENTS.md; or when a platform-specific subagent command is needed for Antigravity or Codex. |
Subagent Launcher
Purpose
Use this skill from the main agent to start the appropriate configured subagent without hard-coding agent-specific instructions in the launcher.
The source of truth for available subagents is AGENTS.md. Each subagent's concrete instructions, required skills, and role live in .agents/agents/<name>/agent.json.
Selection Workflow
- Read the nearest relevant
AGENTS.md.
- Select the subagent whose description and usage match the user's request.
- Read that subagent's
.agents/agents/<name>/agent.json.
- Resolve every skill listed in
agent.json.skills.
- Pass the user's exact task plus any explicitly provided URLs, credentials, files, or constraints.
- Do not use tools directly when
AGENTS.md says the subagent owns that workflow.
Required Skill Loading
Treat agent.json.skills as required runtime dependencies, not as proof that the platform has auto-loaded those skills.
For every skill listed in .agents/agents/<name>/agent.json:
- Resolve the skill path as
.agents/skills/<skill-name>/SKILL.md.
- Attach the skill to the subagent as a platform-native skill item when supported.
- If native skill attachment is unavailable, explicitly instruct the subagent to read and follow that
SKILL.md before acting.
- Do not make the main agent read full skill bodies unless the platform requires the content to be embedded in the subagent prompt.
Antigravity
When define_subagent and invoke_subagent are available:
- Define a subagent using values from
.agents/agents/<name>/agent.json.
- Use full permissions unless the agent config says otherwise:
enable_write_tools: true
enable_mcp_tools: true
enable_subagent_tools: true
- Build the system prompt from:
- The agent config
instructions.
- The resolved skills listed in the agent config, attached by path when supported.
- Any platform-required skill content only if Antigravity cannot attach skills by path.
- Invoke the subagent with the user's task.
Codex
When multi_agent_v1.spawn_agent is available:
- Spawn a
worker agent for execution tasks, or an explorer agent for bounded codebase questions.
- Prefer
fork_context: false unless the subagent needs the current conversation history.
- Pass the selected agent config and required skills as structured items when supported:
{
"agent_type": "worker",
"fork_context": false,
"items": [
{
"type": "text",
"text": "Use this agent config: .agents/agents/<name>/agent.json"
},
{
"type": "skill",
"name": "<skill-name>",
"path": "/home/eliotran/code/.agents/skills/<skill-name>/SKILL.md"
},
{
"type": "text",
"text": "Execute this delegated task: <USER_REQUEST>"
}
]
}
- Repeat the skill item for every skill listed in
agent.json.skills.
- If skill items are unavailable, write a concise prompt that lists every resolved skill path and tells the worker to read and follow them before acting.
- Wait for the subagent only when its result is needed for the next response or next action.
- Close completed agents when they are no longer needed.
Reporting
Report which subagent was used, whether it completed or was blocked, and the result needed by the user. Do not claim success until the subagent reports completion.