multi-agent-orchestration - SKILL.md Agent Skill

name: multi-agent-orchestration description: > Design multi-agent systems using Superagent's supervisor, sequential, and parallel agent types. TRIGGER when: building orchestration agents, routing tasks between sub-agents, designing fan-out/fan-in pipelines, or when the user asks "multi-agent 怎么设计", "supervisor agent", "parallel agents", "agent orchestration". DO NOT TRIGGER when: working on single-model agents (use agent-yaml-authoring) or workflow DAG nodes (use workflow-dag skill). origin: learned tags: [multi-agent, orchestration, supervisor, sequential, parallel, plan-execute]

Multi-Agent Orchestration

Superagent supports four orchestration patterns beyond the single chat_model_agent.

Two-Pass Build Rule

Critical: Leaf agents must be defined (and registered) before orchestration agents that reference them. In configs/agents/, name files so leaves sort first, or use separate files and ensure the runtime loads them in dependency order.

1. Supervisor (Dynamic Routing)

The supervisor LLM decides at runtime which sub-agent handles each message.

spec:
  type: supervisor
  model: gpt-4o
  system_prompt: |
    Route tasks to the appropriate specialist agent.
    - For code questions: use code-assistant
    - For research questions: use research-agent
  agents:
    - name: code-assistant
    - name: research-agent

Agent switch events are emitted as {"type":"agent_switch","from":"supervisor","to":"research-agent"}.

2. Sequential (Pipeline)

Each sub-agent runs in order; output is piped to the next.

spec:
  type: sequential
  agents:
    - name: extractor      # Step 1
    - name: summarizer     # Step 2: receives extractor output
    - name: formatter      # Step 3: receives summarizer output

3. Parallel (Fan-Out / Fan-In)

All sub-agents run simultaneously; results are merged.

spec:
  type: parallel
  agents:
    - name: web-searcher
    - name: db-researcher
    - name: doc-reader
  merge_strategy: concat   # concat | first | vote

merge_strategy:

concat — concatenate all outputs (default)
first — return the first successful output
vote — majority-vote across outputs (useful for fact-checking)

4. Plan-Execute

A planner agent generates a task list; an executor agent runs each task.

spec:
  type: plan_execute
  planner:
    name: task-planner
  executor:
    name: task-executor
  max_iterations: 5

Skill References Between Agents

Any agent can be used as a tool by another agent:

# In the orchestrator's tools list:
tools:
  - skill://research-agent
  - skill://code-assistant

Common Patterns

Research + Write pipeline (sequential):

research-agent → draft-writer → editor-agent

Multi-source research (parallel):

supervisor → [web-searcher, arxiv-searcher, db-searcher] → merge → synthesizer

Approval gate (supervisor + interrupt):

planner → supervisor routes to: executor | approver(interrupt)

Debugging

# Trace agent switches in streaming output
grep "agent_switch" logs/app.log

# Check sub-agent registration order
grep "agent registered" logs/app.log | head -20