Get the eBook free when you register your print book at Manning. Agents turn LLMs into autonomous tools capable of executing on tasks and plans. Multi-agent systems use protocols like MCP and A2A to upgrade the power of a single AI agent with a collaborative AI team. In this book you'll learn how to construct one of these dynamic, powerful, and effective systems from the ground up.
Effectively implementing a multi-agent system requires in-depth infrastructure--and that's exactly what you'll build in this book! Instead of relying on frameworks, you'll design the foundations yourself: the agent loop, tool orchestration, memory, and human-in-the-loop enhancements. Soon, you'll have an in-depth understanding of how multi-agent systems work because you've built your very own!
In
Build a Multi-Agent System (From Scratch) you will learn how to:
- Build a complete LLM agent infrastructure from scratch, including interfaces, tools, data structures, and processing loops
- Orchestrate tool calling with LLMs and connect agents to the Model Context Protocol (MCP) ecosystem
- Implement human-in-the-loop patterns and add memory modules to share state across tasks
- Evaluate agent and multi-agent performance on real tasks
- Add Agent2Agent compatibility so multiple agents can collaborate and solve distributed problems
About the book Build a Multi-Agent System (From Scratch) shows you how to build a complete, working system of agents. Each chapter builds a new stage of your system. Begin by developing your first scratch-built agent, continue to integrate MCP compatibility, incorporate key patterns and designs like human-in-the-loop and memory, and finally implement full Agent2Agent capability that distributes a task among multiple agents. Every milestone comes with a careful walkthrough of the design decisions and code. By the end of the book, you will have a practical, extensible multi-agent system--and the skills to adapt it for research, business automation, or your own experiments.
About the reader For software engineers and AI scientists who know Python, and are familiar with working with LLMs. No specialist hardware required--everything in this book should run on a laptop.
About the author Val Andrei Fajardo is a freelance AI engineer and scientist specializing in LLM agents and AI infrastructure. He is a former founding engineer at LlamaIndex, where he contributed to and maintained their popular open-source Python framework that receives millions of downloads per month. After LlamaIndex, he worked as a researcher at the Vector Institute for AI, where he developed FedRAG, an open-source library for federated fine-tuning of RAG systems, which was accepted into the CODEML workshop at ICML 2025. Andrei holds a PhD in Statistics and Applied Probability from the University of Waterloo.
