Agent-Skills.md

Agent Skills: spawn

How to spawn subagents correctly. Use this reference when you need to delegate work to a specialized agent.

UncategorizedID: c-daly/agent-swarm/spawn

Repository

c-daly/agent-swarm
c-daly

Install this agent skill to your local

pnpm dlx add-skill https://github.com/c-daly/agent-swarm/tree/HEAD/skills/spawn

Skill Files

Browse the full folder contents for spawn.

Download Skill

Loading file tree…

skills/spawn/SKILL.md

Skill Metadata

Name
spawn
Description
How to spawn subagents correctly under the router. Use when delegating work to a specialized agent.

How to Spawn (REQUIRED)

Every subagent runs under the router: its only tool is Bash(mcp*), so it must call tools via mcp-call --caller-id=<its id>. Two steps, never skip either:

  1. Register the agent to get its id + role/workflow briefing: 
    reg = register_agent(agent_id="<task-id>", agent_type="<role>", workflow_id="<workflow>")
  2. Prepend the briefing to the prompt and spawn with the role as subagent_type: 
    Task(
  subagent_type="<role>",
  prompt=reg["briefing"] + "\n\n" + <task description + working dir>,
)

reg["briefing"] carries the agent's id, the mcp-call/--caller-id convention, its role rules, and its workflow phases. Without it the agent has no identity and its tool calls are denied.

Do NOT spawn native subagent types (Explore / general-purpose / Plan). They use Claude Code's built-in tools, which the router blocks — they cannot reach mcp-call correctly and will flail.

Agent Selection

agent_type and subagent_type are the same role name.

| Need | Role | Model | |------|------|-------| | Find code/files | explorer | haiku | | Web/doc research | researcher | haiku | | Plan changes | architect | sonnet | | Write code | implementer | sonnet | | Review changes | reviewer | sonnet | | Fix bugs | debugger | sonnet | | Git operations | git-agent | haiku |

Prompt Structure (after the briefing)

  1. Task -- one clear objective
  2. Scope -- exclusive file list (no other agent touches these)
  3. Output -- specific return format
  4. Constraints -- what NOT to do
  5. Acceptance criteria -- machine-checkable (grep, test exit codes)

Never embed file content in prompts. Give intent + verification criteria.

Model Selection

  • Can downgrade (sonnet -> haiku). Never upgrade without asking the orchestrator.
  • Default to cheaper when unsure.

Parallel Spawning

Independent tasks -> register + Task each, multiple in ONE message block. Max 5 parallel agents.

File Ownership

Each file modified by AT MOST one concurrent agent. Overlap -> serialize with depends_on, or merge into one task.

Anti-Patterns

  • Don't spawn native subagent types (Explore/general-purpose/Plan) -- no router access
  • Don't spawn without registering and prepending the briefing
  • Don't spawn for one-liner tasks or <3 tool calls
  • Don't use opus for subagents (orchestrator only)
  • Don't give vague prompts ("look around")
  • Batch related work into one agent