Preface

Why write this in the open

This book is being written in public because the subject will not sit still long enough for a print schedule. A model release changes what is possible in a fortnight. A regulation changes what is permitted in a quarter. Anything written to a twelve month publishing cycle is a historical document by the time it lands.

So the manuscript lives here, as HTML, with a commit hash at the top of the page. When a chapter changes you can see when it changed. When I am wrong about something you can watch me correct it. That is a weaker promise than a finished book makes, and a more honest one.

The PDF and EPUB are built from the same source on every push. They are always the current state of the argument, not a snapshot taken months ago.

Who this is for

This is written for engineers who are already good at the job, and who suspect that being good at the job is about to mean something different. Not for people looking for prompt tricks. There are enough of those.

• Staff and principal engineers deciding what to standardise on
• Platform and DevEx teams who will own the harnesses everyone else uses
• Engineering leaders being asked for a number on productivity, who know the number is not real
• Architects working inside regulated European enterprises, where provenance is not optional

You do not need to have used an agent harness in anger. You do need to be willing to accept that the part of the job you are proudest of might be the part that automates first.

What the book claims

Generation stopped being the constraint. Specification became the constraint. Almost everything else follows from that one move.

Three claims run through every chapter. First, that the specification is now the primary engineering artefact, and code is a derived output. Second, that verification has to be independent of generation or it verifies nothing. Third, that provenance is the price of admission for any of this inside a regulated organisation.

If you disagree with the first claim, the rest of the book will read as an elaborate mistake. That is a reasonable place to stop.