Git took years to settle into something we all agree on, so did Scrum. And they settled the same way every time, people writing down what they actually did, arguing about it in public, and slowly converging. Agentic coding is going through the same thing now, only faster. But the way we work with agents should be built like that too, evidence based and together, not by whoever has the strongest opinion this week.

So I wrote a book. Well, I commissioned a book. The Agentic Playbook is a field guide with twenty plays, small named moves like “Starve the context”, “Make the agent prove it” and “Know when not to use an agent”. Every play has the same five headings (the problem, the play, a worked example, the way it goes wrong, a checklist), so one can open any of them and try it that afternoon.

It is my piece of the writing-down part. Every figure in the book is dated and traced to a cited research note in the repo, so one can check it and argue with it. And the evidence is thinner than one would think. Nearly every published study of “AI coding productivity” measures autocomplete or chat, not agents, and several of the most quoted figures do not survive being looked up. So there is a whole part, Where It Struggles, that says so.

How was it written then? With agents. Several of them, in parallel, on separate chapters, against a shared plan, a style guide and a pile of research notes. And the neat part is that the repo had to follow its own advice to make that work. The CLAUDE.md is short and points to context cards, loaded only when the situation matches:

- [book-structure](cards/book-structure.md) — writing or editing any part of the book itself
- [research-notes](cards/research-notes.md) — running a research pass, or producing anything
  that a later writing task will cite

That is play number five, “Split the agent file into cards”, applied to the book that describes it. Most of the failure modes in there were not researched from a distance, they turned up in the drafting. Me smile.

For reviewing there is a little server that lets one annotate the text in the browser and hand the notes back to the agents:

make review   # http://127.0.0.1:8777

It is about 42,000 words, CC BY 4.0, and readable in one page. If that is too much, there is a short presentation with the overview.

Grab the source at github.com/epatel/agentic-playbook, and if one of the plays is wrong, feel free to tell me. We are supposed to figure this out together after all…