Build a Multi-Agent System (from Scratch)

Build a Multi-Agent System (from Scratch) is a Manning Early Access Program (MEAP) book by Val Andrei Fajardo, a founding engineer at LlamaIndex who later researched federated RAG systems at the Vector Institute and holds a PhD in Statistics and Applied Probability from the University of Waterloo. The MEAP began in October 2025; seven of nine chapters are currently released and publication is estimated for November 2026, at roughly 325 pages. Rather than teaching LangGraph or CrewAI, the book builds its own framework so the internals are visible. Part 1, Our First LLM Agent, covers what LLM agents and multi-agent systems are, working with tools, working with LLMs, and the LLM agent class. Part 2, LLM Agent Enhancements, adds MCP tools, Agent Skills, memory and human-in-the-loop. Part 3 builds multi-agent systems with the Agent2Agent (A2A) protocol, and an appendix implements a Pydantic-based function tool. The companion repository, nerdai/llm-agents-from-scratch (Apache 2.0, about 186 stars and 577 commits), is ahead of the published outline: it already splits Part 3 into a subagents chapter and an A2A chapter, ships notebooks for each chapter, and includes a Monte Carlo pi estimation capstone. Code runs locally against Ollama with qwen3:14b by default, which needs about 16 GB of RAM and ideally a GPU, and a Lightning AI template is offered for readers without that hardware. Because it is still in early access, expect revisions and two chapters still to arrive.

Topic

Agentic

Level

Intermediate

Format

Book

Cost

Paid

Duration

~325 pages (estimated), 9 chapters plus per-chapter notebooks, self-paced

Provider

Manning Publications (Val Andrei Fajardo)

Hands-on

Yes — code/exercises

Certificate

None

• ✓Implement a tool interface and orchestrate LLM tool calling without an existing agent framework
• ✓Write an LLM abstraction and a core agent class that runs the processing loop
• ✓Connect a scratch-built agent to the Model Context Protocol ecosystem through MCP tools
• ✓Add Agent Skills support so agents can reuse packaged, repeatable workflows
• ✓Build memory modules that share state across an agent's tasks
• ✓Implement human-in-the-loop patterns that bring a person into an agent's task execution
• ✓Distribute one task across multiple cooperating agents using the Agent2Agent (A2A) protocol
• ✓Evaluate single-agent and multi-agent performance on real tasks

• •Written by a founding LlamaIndex engineer, so the from-scratch design is informed by building a production agent framework
• •Covers the current agent interoperability stack in one codebase: MCP, Agent Skills and A2A
• •Actively developed Apache 2.0 companion repository with per-chapter notebooks, 577 commits and a capstone project
• •Examples run locally on Ollama with an open-weight Qwen3 model, so they do not depend on a paid model API
• •Early-access purchase delivers chapters as they land, ahead of the estimated November 2026 publication

This page was written from 3 sources, 1 on domains other than manning.com.

1. 1.manning.com — build a multi agent system from scratchvendor
2. 2.manning.com — chapter 1vendor
3. 3.github.com — llm agents from scratch