I built a book. It is called *From Copilot to Colleague: How AI Engineering Turns Models into Dependable Systems*. You can read it online at fromcopilottocolleague.com.
It is also a public experiment. The question was whether autonomous research loops, working from a real corpus, can produce a source-anchored book on AI engineering. The loops drafted the manuscript. The repository holds the method that produced it.
The book has one thesis. Models create possibility. Scaffolding creates trust. Organizations decide whether that trust compounds. The ten chapters are the argument for that line.
What the book covers
The book has ten chapters in four acts. The first act, chapters 1 and 2, states the problem. AI is moving from answering questions to doing delegated work, and cheap generation makes human judgment more visible, not less.
The second act, chapters 3 to 7, builds the scaffolding stack. It covers harnesses and specs that agents can use, evals as a control system, context as infrastructure, runtimes with a human control plane, and security and identity for systems that act.
The third act, chapter 8, is the stress test. Realtime voice removes the hiding places that text chat gives a weak system. The last act, chapters 9 and 10, widens the view. It asks what an AI-native organization looks like and which ideas endure when the tools change.
Where the material comes from
An AI-drafted book with no source of truth only remixes what the model already knows. I wanted a source of truth. The corpus is 794 talks from practitioners on the AI Engineer YouTube channel.
The rule of the project is simple: no anchor, no claim. Every claim in the manuscript carries a video, a start and end time, and a verbatim quote. The website shows these under each chapter, so a reader can jump to the exact moment in the talk.
How the quality is checked
No single model should grade prose shaped like its own output. A panel of three AI models from different families scores each chapter on six dimensions: humanness, voice, usefulness, evidence density, claim defensibility, and redundancy. The median score is the verdict. When the three models disagree by more than 20 points, that cell is flagged as the place to look.
The repository holds the whole method: the agent instructions, the research passes, and the judges. A bounded research loop takes one mission at a time, commits its result, and logs the pass. Every step is reversible and sits in git.
What I will post here
The book is the proof that the method works. The method is the deliverable. Over the next posts I will go through the book chapter by chapter. Each chapter gets three short posts: its core argument, a practical way to apply it this week, and one scene or example that teaches a single lesson.
Start where your own work hurts. If agents disappoint you, read about harnesses. If you cannot tell whether a change made things better, read about evals. If you ship to real users, read about trust.
This post draws on the whole of From Copilot to Colleague.