The Orchestrator

Some of you are reading this and thinking: this isn’t here yet. You are right, and I would rather concede it in the introduction than have you catch me at it in chapter seven. In most organizations we are still building the tools that will make this job real. Systems that can reliably inspect a codebase, propose a change, run the tests, evaluate the result, and prepare a release without a person correcting them at every step are not yet widely available, and the persistent runtimes, evaluation frameworks, governance layers, and permission envelopes are mostly early, rough, or missing altogether.

The pieces are visible, though, and some of them are further along than the skepticism allows. Andrej Karpathy — a founding member of OpenAI who went on to direct AI at Tesla — said in a March 2026 interview that he had probably not typed a line of code since December, and that his split between writing code and delegating to agents had flipped from roughly 80/20 to 20/80 and kept going.¹⁴ JPMorgan Chase reported two hundred thousand employees on its internal LLM Suite within the first eight months; by December 2025 its chief analytics officer described two hundred and fifty thousand users — more than 60 percent of the firm — building and customizing their own assistants, and argued that the models themselves are commodities while the durable advantage is the connectivity around them.¹⁵ The work is forming well ahead of the name.

Writing about that gap requires a method, and the one I use here I call near-future nonfiction. The term describes work that has appeared in pieces without consolidating into a recognized profession. The evidence is real — the companies, the tools, and the people are real, and the claims above are attributed to the sources that reported them — but no single person or organization yet represents the full scope of what this book describes. So when the book uses a short fictional narrative to show what a day of orchestration looks like, it is not claiming that person exists. It is assembling a picture out of parts that already do.

I did not invent the phrase, and the way I arrived at it is more interesting than the phrase itself. I recently asked Julian Bleecker — whose work on design fiction shaped how I think about prototype futures — whether “near-future nonfiction” was a legitimate description of what I was doing. His response was roughly: wait, isn’t that just about everything I’m doing?¹⁶ The label was already familiar to him. What was new was my route to it. I wrote the fiction first, in the Condition Set trilogy, and found that the world-building kept pointing me back at a reality arriving faster than most people had noticed. The novels were never the evidence. They were the exercise that forced the questions. The evidence is in the tools, the companies, and the work already underway, and near-future nonfiction is what I am calling writing that treats those unfinished pieces as reportable without pretending the profession has consolidated.

Klarna is the case most often held up as what it looks like when the orchestration is missing, and it deserves more care than it usually gets. In February 2024 the Swedish fintech announced that its AI assistant, one month live, had handled two-thirds of its customer-service chats and was doing work equivalent to seven hundred full-time agents, with company-reported satisfaction on par with human agents. Equivalent workload is not the same claim as seven hundred people replaced, and the numbers are company-reported rather than independently measured. What later reporting and the company’s own materials show is a renewed emphasis on human service — including an on-demand human-workforce experiment — not an abandonment of AI support.¹⁷ The lesson I draw is not that Klarna failed and reversed. It is that task-equivalent capacity and workforce substitution are different claims, that escalation paths and quality evidence are exactly what an announcement like that should disclose, and that an organization that cannot tell the two apart is already mis-orchestrating. And the asymmetry is worth noticing: getting this right means things work roughly as expected, while getting it wrong means real people are on the receiving end of decisions nobody designed.

I wrote this book because we are at the beginning of something, and beginnings are the only point at which the shape of a thing is still negotiable. The pages that follow describe a job and a practice that do not fully exist yet. But the parts keep showing up — in coding agents, in persistent runtimes, in skills and memory systems, in multi-agent frameworks, and in the daily experience of people already doing this work without a name for it. What you read here is going to turn into something real. It may not look much like what is on these pages; the names will change and the tools certainly will. My forecast is that the pattern will not — and the evidence for that forecast is everything in the preceding pages. A person who designs and governs a system of delegated intelligence, who holds both the authority and the accountability, and who decides when the system may continue and when it must stop — that pattern is already here.

This book is not a description of the present. It is a recognition of the beginning.