Preface

Name: AI Agent Architecture: From Single Agent to Enterprise Multi-Agent Systems
Author: Wayland Zhang

Why I Wrote This Book

In the second half of 2025, I shifted from explaining AI theory to engineering Agent implementations. At first, I thought building an enterprise-ready Agent system using open-source libraries would be straightforward. Then when I actually started building a production-grade Agent system, I ran into a problem: tutorials on the market either stopped at the "call an API to make a chatbot" demo stage, or were just translations of some framework's documentation.

When I needed to answer these questions, I couldn't find answers:

How do multiple Agents collaborate? DAG or Supervisor?
How do you allocate token budgets? Per call or across the whole workflow?
What happens when tool execution fails? How do you persist state?
How do you do access control and auditing in enterprise environments?

These questions had no ready-made answers -- I could only figure them out by trial and error. Since frameworks on the market all had various shortcomings that couldn't meet enterprise-grade requirements, I decided to write one from scratch. Shannon is the product of that exploration -- a three-tier architecture multi-Agent system implemented in Go/Rust/Python.

This book is the systematic organization of those hard-won lessons.

The Writing Philosophy of This Book

Patterns first, frameworks second.

Most Agent tutorials on the market are tied to specific frameworks -- how to use LangChain, how to configure CrewAI. Frameworks become outdated, but patterns don't.

Every chapter in this book follows this structure:

Start with the problem -- What scenario needs this capability?
Then the pattern -- What's the universal design pattern? (framework-agnostic)
Then the implementation -- Show one implementation approach using Shannon as an example
Finally, comparison -- How do other frameworks solve the same problem

If after reading a chapter you can implement the same pattern in any framework -- that chapter has succeeded.

Who This Book Is For

This book is for you if:

You want to build production-grade Agent systems, not just demos
You need to handle multi-Agent collaboration in complex scenarios
You care about cost control, security, observability -- enterprise-grade concerns
You want to understand the design patterns behind various Agent frameworks
You're a backend developer, architect, or technical lead

This book may not be for you if:

You just want to quickly call the ChatGPT API (just read the official docs)
You need a collection of Prompt Engineering tricks (there are more specialized resources)
You've never encountered LLM-related concepts (recommend learning basics first)

How to Read This Book

Quick Start (2-3 days):

All of Part 1 -> Chapter 3 -> Chapter 13 -> Chapter 20

Goal: Understand Agent basics, tool calling, multi-Agent orchestration, production architecture

Systematic Learning (2-3 weeks):

Read Part 1-8 sequentially, with Shannon code alongside

Goal: Complete mastery from single Agent to enterprise multi-Agent

Hot Topics Track (1-2 days):

Chapter 4 (MCP) -> Chapter 27 (Computer Use) -> Chapter 28 (Agentic Coding)

Goal: Learn about the hottest Agent topics of 2025-2026

Every chapter ends with a "Chapter Recap" that you can use to verify whether you truly understood the core concepts.

About the Code

This book uses Shannon as the reference implementation, but it is not a Shannon user manual.

Shannon uses a three-tier architecture:

Orchestrator (Go)    - Orchestration, budget, policy
Agent Core (Rust)    - Execution, sandbox, rate limiting
LLM Service (Python) - Inference, tools, vectors

The code examples in this book demonstrate design patterns, not framework APIs. You can absolutely implement the same patterns using LangGraph, CrewAI, or your own framework.

About Timeliness

The AI Agent field is evolving extremely fast. MCP protocol was only released at the end of 2024, and Computer Use is still rapidly iterating.

This book clearly marks highly volatile content:

Timeliness Note (2026-01): This section's content is based on MCP specification v1.0. Please check the latest documentation to confirm any updates.

Core architectural patterns are relatively stable, but specific APIs and tools may change. When in doubt, defer to official documentation.

Acknowledgments

Thanks to all colleagues and friends who helped during the Agent system building process.

Thanks to open-source community contributors -- LangChain, LangGraph, CrewAI, AutoGen and other projects let us stand on the shoulders of giants.

Thanks to Claude Code, OpenAI Codex programming tools and models for pushing the boundaries of rapid code writing, turning this Agent orchestration framework and platform from concept into reality quickly.

Wayland Zhang

January 2026

"The best way to learn is to build."

Errata and Feedback

If you find errors in the book, or have any suggestions, please contact us through:

GitHub Issues: This book's repository
Shannon OSS: github.com/Kocoro-lab/Shannon

Technical books inevitably have omissions -- thank you for your understanding and help.