Type: url
Original: https://jimmysong.io/zh/book/ai-handbook/agent/multi-agent/
Status: generated
Summary
A draft chapter of Jimmy Song’s 智能体构建指南 on multi-agent collaboration, written as an operations handbook rather than a concept survey. It covers role separation (Planner, Worker, Reviewer, Orchestrator), consensus and arbitration (majority voting, confidence-weighted voting, trust-value routing), task-plan merging and conflict resolution, cross-agent tracing through shared state machines and trace IDs, and resource containment through call quotas, budget monitoring, retry limits and circuit breaking. It then compares six coordination frameworks — LangGraph, CrewAI, Agno, Mastra, Google ADK and AWS Strands — against centralized, hierarchical, hybrid and decentralized architectures, argues that none of them solves selective semantic context transfer between agents, and records a 2026 consolidation of the ecosystem in which MCP connects tools and A2A connects agents.