Practical Multi-Agent AI Systems: How to Architect, Build, and Scale Next-Generation AI Systems That Work in the Real World
暫譯: 實用的多代理人工智慧系統：如何架構、建設和擴展能在現實世界中運作的下一代人工智慧系統

A practical guide to building multi-agent AI systems that earn trust in production

In this hands-on guide to production agentic AI, Murali Kashaboina explains how to design, build, and operate hierarchical multi-agent systems that are reliable enough for regulated enterprise environments. He demonstrates the complete lifecycle from choosing the right architecture pattern through deployment to the operational concerns that surface on day two, using an example customer service platform that handles real queries against a real Knowledge Graph with real security controls.

Practical Multi-Agent AI Systems walks you through building a system with five specialist agent teams, 20+ worker agents, and multiple agentic tools, orchestrated by LangGraph with MCP tool servers and A2A agent-to-agent delegation. But unlike resources that stop at implementation, this book devotes equal attention to what happens after deployment. You'll learn how to trace a wrong answer across five agents using a Context Graph integrated with LangFuse observability, how to classify failures into actionable categories, and how to build feedback loops that improve agent behavior using few-shot examples from production without fine-tuning the LLM. You'll also find targeted coverage of the operational concerns most books skip entirely: context compression across multi-turn conversations, provider failover with circuit breakers, token budget enforcement, prompt injection through tool outputs, and a systematic evaluation framework using DeepEval that makes non-deterministic systems testable.

Perfect for AI engineers, solution architects, and engineering leaders building agentic systems where accuracy, auditability, and security are non-negotiable, this book is an indispensable resource for every practitioner who wants to learn how to:

• Select and implement multi-agent architecture patterns with clear trade-off analysis
• Engineer context across agent boundaries using a seven-category taxonomy
• Enforce production security with mTLS, encryption, RBAC, guardrails, and PII scrubbing
• Build unified observability that shows the actual LLM prompt and response at every decision point
• Diagnose and fix failures using a structured taxonomy and inline trace cards
• Test non-deterministic systems with statistical evaluation and measurable quality baselines