Production-Grade Multi-Agent Systems: From Architecture Decisions to Engineering Delivery
A Chinese-language open-source book for AI engineers and architects on building production multi-agent systems, covering tool calling, coordination, context engineering, safety, evaluation and delivery across ten chapters.
This open-source book, published in Simplified Chinese, English and Japanese, works from the question of which part of a system genuinely needs to decide its own next step from environment feedback. Its ten chapters cover controlled tool calling, multi-agent collaboration, context engineering, production infrastructure, security, system integration, evaluation, AgentOps and verifiable delivery.
Alongside the text, the repository provides an independent reference implementation built around an insurance-claims investigation scenario that puts the book's tool contracts, MCP usage, agent-to-agent protocol, context graph, tool guarding, evaluation datasets and release evidence into a single runnable system with locked dependencies, non-root containers, and security and evaluation gates. The book also ships 68 original architecture diagrams and 10 reusable engineering templates, and can be read locally with MkDocs after installing its Python requirements.