Most material about AI agents describes what is possible. This one describes what breaks.

Seventy-five pages on the part nobody puts in a demo: the deterministic software wrapped around the model. When the loop is allowed to stop. What is still in the context window on turn six. Which claims get checked before they are sent. What the agent may never decide on its own.

Nine parts, six worksheets, three diagrams. A six-clinic dental front desk runs through the book as a worked example, with shorter cases from a bakery, a freight broker, a parts distributor and a law firm. Each part exists because of a failure that only turns up once real traffic arrives.

Inside: the three eras of leverage. The agent loop, with stopping rules, cost per turn and context rot. The five components of a harness, and the symptom you get when each one is missing. Reaching real systems without causing a security incident. Packaging workflows your team can version and test. The three training stages, and which problems each can actually fix. Evaluation that grades a system under stress. Ten boundaries in a working stack. Five builds worth doing, and a ninety-day plan.

Written for engineers shipping their first agent, product managers who have to review one, and founders deciding what to build in house. You need to have used a chat model. You do not need to have trained anything.

Print-ready A4 PDF.