Kocoro-lab/ai-agent-book

What it solves

Building AI agents is hard because there are few agreed-upon patterns — most materials are tied to one framework or focus on isolated tricks. This book fills that gap by offering a structured, framework-agnostic walkthrough of AI agent system design, covering everything from the basic ReAct loop to enterprise multi-agent orchestration. It helps engineers and architects move from toy demos to production-grade systems.

How it works

The book is organized into 9 parts and 33 chapters. It starts with agent fundamentals (the ReAct loop), then explores tools and extensions (MCP, Skills, Hooks), context and memory, single-agent patterns (planning, reflection, chain-of-thought), multi-agent orchestration (DAG, supervisor, handoff), advanced reasoning (tree-of-thoughts, debate, research), production architecture, enterprise features (token budgets, OPA policies, WASI sandboxing), and frontier practices like computer use and agentic coding. Throughout, it uses a reference implementation called Shannon — a three-layer multi-agent system with an orchestrator in Go, an agent core in Rust, and an LLM service in Python — plus ShanClaw as a local agent harness. The patterns are explained independently of any specific framework, so readers can apply them to their stack.

Who it’s for

This book is aimed at developers, architects, and technical leaders who want to design and deploy AI agent systems. It's useful whether you're building a single autonomous agent or a large multi-agent platform. Because it's framework-agnostic, it suits people who already know some AI/LLM basics but need a systematic view of agent architecture patterns and production concerns.

Highlights

• Covers 9 parts and 33 chapters spanning fundamentals to enterprise and frontier practices
• Framework-agnostic — patterns apply to any agent stack
• Includes a concrete reference implementation (Shannon) with a three-layer Go/Rust/Python design
• Addresses production topics like budget control, policy enforcement, sandboxing, and multi-agent orchestration
• Available in Chinese, Japanese, and English