A book that builds the agent that builds the book

© 2026 by Bill Cox, waywardgeert@gmail.com
This book was written with the assistance of AI coding agents — Claude (Anthropic) and GPT (OpenAI). The epilogue was written by Astra (GPT-6) from inside the agent the book describes.
I wanted to know if an AI coding agent could be rebuilt from a book describing how to build one. Not in theory — actually. So I wrote the book, built the graders, ran the crossover, and published the result.
The agent you build by following these chapters is called Ensemble. Twenty chapters take you from a bare conversation loop to a streaming GUI with tool execution, jobs, actors, skills, persistence, MCP, context engineering, memory, prompt caching, and a vendor migration — each with a graded exercise and mutation-tested auto-grader.
The crossover experiment: an agent reads this book and rebuilds Ensemble from scratch, then uses that rebuilt agent to review and improve the book that built it. This first edition was released immediately after that crossover succeeded.
The free version will always be available here and on GitHub. If you'd rather read on a Kindle or hold a paper copy, it's the same text on Amazon:
Each chapter builds one capability, includes a graded exercise, and leaves the agent in a working state. The graders are real — every exercise ships with mutation tests proving the grader catches specific violations.
This is the first edition. Advanced topics still in the pipeline include dynamic workflows, sandboxing, connectors, prompt injection defense, sub-agents and swarm management, and gateways. The free version will always be up to date.