“Done” is a claim, not a fact. The deploy reports success. The test suite goes green. The agent marks the task complete and closes it — cheerfully, confidently, and sometimes while the thing it built isn’t actually there. The first book was about not trusting the model when it says it worked. This one is the same distrust turned toward building: what happens when you try to make genuinely useful things with a fleet of AI agents, and they tell you it’s finished when it isn’t.
It is a practical book — real systems that are still running because switching them off would hurt — and its through-line is the one skill that separates a working build from a demo: refusing to believe “done” until you have watched it run with your own eyes. The sober audit. Verifying the deployed thing and not the repo. A second pair of eyes that isn’t you. Making the machine refuse the lie. The one button you never let it press.
If you build things with these tools — or you are about to — this is the book about the gap between “it works” and it working. Most of what goes wrong lives in that gap. So does most of the interesting engineering.
Listen to the whole thing, free
The complete book, narrated by Hugh Merrick — the same voice as the first book, 1 hour 30 minutes across 14 chapters. Listen to a sample below, then download it to keep. DRM-free, no signup, no payment. It’s here to be read and heard, not sold.
Sample
Two formats: a single .m4b for audiobook apps (Apple Books, BookPlayer, VLC) that remembers your place and carries chapter markers, or an MP3 zip of the chapters that plays on anything, including Android and desktop.
Steve Parker is a software consultant with thirty-seven years in enterprise systems who now runs twenty-five AI agents from a spare room and a Raspberry Pi. He is as surprised as you are.