AgentSkillsCN

multi-agent-orchestration

协调多智能体的模式设计。可用于为复杂工作流构建层级式、菱形式、点对点式，或协作式架构。

SKILL.md
--- frontmatter
name: multi-agent-orchestration
description: design patterns for coordinating multiple agents. Use this to implement Hierarchical, Diamond, Peer-to-Peer, or Collaborative architectures for complex workflows.

Multi-Agent Orchestration Patterns

Goal

Design robust architectures where specialized agents collaborate to solve complex problems that are too difficult for a single agent.

Core Design Patterns

1. Hierarchical (The "Manager")

  • •Structure: A central Orchestrator Agent acts as a router. It classifies the user's intent and delegates the task to a specific specialized agent (e.g., Navigation Agent, Media Agent).
  • •Usage: Best for tasks with clear, distinct domains where a single specialist can handle the request.
  • •Flow: User -> Orchestrator -> Specialist -> User.

2. The Diamond Pattern (The "Moderator")

  • •Structure: Similar to Hierarchical, but the specialist's output doesn't go back to the user immediately. It passes through a second central node (e.g., a Rephraser or Safety agent) to standardize or polish the final response.
  • •Usage: Essential when you need consistent tone, style, or safety guardrails across diverse specialists.
  • •Flow: User -> Orchestrator -> Specialist -> Rephraser -> User.

3. Peer-to-Peer (The "Network")

  • •Structure: Agents can hand off tasks directly to one another without returning to a central orchestrator.
  • •Usage: Best for resilience. If the Navigation Agent realizes a query is actually about history, it can hand off directly to the Knowledge Agent, recovering from initial routing errors.
  • •Flow: User -> Agent A -> Agent B -> User.

4. Collaborative (The "Mixer")

  • •Structure: Multiple specialists work on the same query in parallel. A Response Mixer agent then takes the partial answers from all of them and synthesizes a single, comprehensive response.
  • •Usage: Best for multifaceted queries (e.g., "How do I handle hydroplaning?") where you need safety data, physics explanations, and driving tips simultaneously.
  • •Flow: User -> [Agent A, Agent B, Agent C] -> Mixer -> User.

Orchestration Best Practices

  • •Specialization: Ensure each agent has a narrow, well-defined domain. Specialized agents are easier to optimize and debug than generalists.
  • •Adaptive Loops: Implement retry loops where an agent can reformulate its own query if the initial attempt fails (e.g., broadening search terms if no results are found).