I’m finishing a book about cognitive sovereignty — The Anti-AI Brain. The book is about how a person stays able to think when machines do more and more of the thinking for them. In parallel, I’m building skills, CLAUDE.md files, and system prompts for our label. Two tabs, two tasks. At some point I noticed it’s one task seen from two sides.
A person with nobody home. An AI stack with nobody home. Same disease. You’re either looking in a mirror or at the code.
What everyone assembled
A standard AI stack in 2026 is no longer “model plus prompt.” Multi-agent setups with handoffs. Short- and long-term memory over a vector store. MCP servers wiring the model into dozens of external systems. Sub-agents delegating to each other. Evals, traces, guardrails, observability. All of it packs into Claude Agent SDK, OpenAI Agents SDK, LangGraph, and installs with one command. The harness — in the industry’s sense of the word — comes together in a day.
And it works. Sometimes it even looks like something genuinely useful.
Six slots out of seven. Context, tools, skills, rights, evals, review — present. The seventh, owner, is empty. Capability is assembled. The thing capability is supposed to serve — nobody assembled. Hence “sometimes.”
Gurdjieff already said this
George Gurdjieff — mystic and philosopher of the early twentieth century, founder of the “Fourth Way” teaching. He has a parable about a chariot that he used to explain how a human being is built. Four parts. The carriage is the body. The horse is the emotions. The driver is the mind. The master is consciousness, the one for whose sake the whole contraption is going anywhere at all.
For most people, Gurdjieff says, the lower three parts are in bad shape. The carriage is rusted. The horse is untrained, bucking. The driver is drunk and asleep on the box. As for the master — either there isn’t one, or he’s somewhere in the back room and doesn’t remember he owns a carriage.
The diagram maps onto the harness too cleanly to ignore. The horse is the model. The carriage is tools. The straps between them are context — CLAUDE.md, skills, system prompts. The driver is evals, review, guardrails. The master is owner. The harness in its industrial sense is the first four parts. Gurdjieff reminds you that without the fifth, the first four go nowhere.
What nobody assembled
Teams build capability like assembling a kit. New Claude, new Cursor, new agent, new MCP server, new pipeline. Tools — there. Prompts — slapped on. Evals — sure, but per LangChain’s “State of Agent Engineering 2026,” only 52% of teams actually run offline evals. The rest do it in “looks fine, ship it” mode.
And off they go. Gartner forecasts that 40%+ of agentic AI projects will be cancelled by 2027. The loud causes are brittle connectors, context rot, compounding errors. That’s engineering work, and it’s necessary. Underneath it lies a second layer: even once the engineering problems are solved, the team still can’t explain what in their AI stack makes their company theirs. Output is generic — could have been written by anyone. Useful, sometimes even smart, never theirs. This layer LangGraph and Agent SDK don’t close, because closing it isn’t code.
Gurdjieff has the exact word for this — sleeping. The circuit is formally assembled, motion is happening, the master isn’t there. The machine moves, the “employee” works, and the team can’t answer a simple question. What should the result look like for you to see it and say “yes, this is ours”?
Writing the harness is work that isn’t about AI
You can’t buy the harness ready-made. You can’t pull it off GitHub.
The harness is text-encoded knowledge of how things ought to be at your shop. If that knowledge doesn’t exist, the file doesn’t close.
A simple example from yesterday. I ran my own 22 installed skills through a checklist — does each one have a full harness contract, or only a name and description. Result: 5 out of 22. The other 17 — just those two fields, everything else implicit.
The most systemic gap: none of the 22 declared what they’re allowed to do. Write externally, spend money, delete files — all of it gated at runtime by the environment, not at the skill-level contract. Owner — missing on 17 of 22.
I found this hole at home. Not at a client, not in a case study. At home, in yesterday’s audit. The exact master this post is about — under-specified across most of my own infrastructure. The sharpest risk on my list: a skill that can create autonomous scheduled tasks, no harness contract. One buggy iteration equals two hundred batches running while I sleep.
AI infrastructure is about whether you’ve woken the master inside your own head.
The same disease, viewed from the other side
The book I’m finishing is about the same disease, viewed from the other side. The Anti-AI Brain is about cognitive sovereignty — what happens to your brain when machines do the thinking, and what to do about it.
The thesis is simple. AI doesn’t make you stupid. It relocates your intelligence to a server. Efim Ostrovsky called this RAI — Remote Access Intelligence — back in 2019. When the server is unreachable, you have no intelligence. You didn’t “use a calculator.” You forgot arithmetic.
The two sides close on each other. If it’s empty inside, you won’t write a harness outside either. The harness is your head, externalized into a file. If the inside is a tour of other people’s best practices, the outside ends up being the same tour.
The sharpest example in the book: Anthropic engineers. Dario Amodei in an interview: “I have engineers at Anthropic who don’t write code. They let Claude write the code and they edit and review.” 50+ major releases in 52 days. These engineers climbed a rung. From Interface to Identity. From execution to architecture. AI does the hands, the human does the head. They could do this because the master wasn’t asleep. They knew what “the right code” looked like when Claude was the one writing it.
At the same Anthropic they keep philosophers on staff. Amanda Askell — PhD in philosophy from NYU, in charge of Claude’s character. Joe Carlsmith, Ben Levinstein, Jackson Kernion — also there. At Google DeepMind, Iason Gabriel — formerly taught moral philosophy at Oxford. Six-figure salaries. From below, engineers climbed from Interface to Identity. From above, philosophers are hired so they can be the master for the model itself. Where code ends, the master begins.
The opposite example — OpenAI. In May 2024, both heads of the alignment team left in the space of three days: Ilya Sutskever and Jan Leike. Leike’s farewell tweet: “safety culture and processes have taken a backseat to shiny products.” The Superalignment team was dissolved the same day. In October 2024, the AGI Readiness team was dissolved. In February 2026, Mission Alignment was dissolved after 16 months. Three teams whose work was the question “what does safe mean” and “what does good mean” — closed in two years. Today OpenAI is losing momentum to Anthropic, losing investor confidence ahead of its IPO, and tied up in an ongoing public trial with Musk. You can’t fire the master three times in a row and get away with it.
Karpathy, earlier in 2026, put it harder: “Intelligence starts to puppet humanity. Humans become its actuators and sensors.” That’s the second caste — the one without a master. Their system doesn’t ask them anything. They become actuators.
The choice — Partisan or actuator — comes up every time you sit down to write a CLAUDE.md or solve a task with AI. Decide it inside your own head: Partisan. Rewrite the first thing the AI gave you: actuator.
How to tell if the master is there
A test. Open your CLAUDE.md, or whatever system file your AI reads first. If it doesn’t exist, the master is asleep. If it exists but was written by AI, the master is sleeping more deeply — he couldn’t even formulate his own rules. He delegated himself.
If the file exists and a human wrote it, read it and ask three questions.
First. Does the text show that a specific human or team wrote this — with their scars, their precedents, “we got burned once and we don’t do it that way anymore”? Or is it generic best practices lifted from any AI blog?
Second. Is there a single line in there that contradicts “what everyone does”? What makes you you almost always looks like a refusal of some convention. If the file contains no refusals, the file isn’t yours.
Third. What happens if the model breaks these rules? Who notices? If the answer is “nobody, until prod falls over” — the driver is asleep. If it’s “nobody, we’ll check once a quarter” — the driver is asleep more deeply.
The three questions measure one thing: whether the work on yourself made it onto paper.
So
The models are fine. They’re smart. Long prompts don’t fix this. Long prompts are an attempt to compensate, in words, for the absence of a master.
AI needs an owner who’s awake. Inside the head — so there’s something to externalize into the harness. Outside, in the harness — so there’s somebody to answer for it. One doesn’t work without the other.
When the master is in place, the harness pulls tight, the horse pulls, the carriage rolls, the driver doesn’t fall asleep. The rest of the parts feel it.
Gurdjieff wrote a hundred years ago that the goal of the teaching is to wake up the master in a human being. Today the same question returns from the other side. Open your CLAUDE.md and see who wrote it. If the answer is “nobody in particular” — you don’t have a master in the file or in your head.
Maybe that’s the main use case for AI: to make visible whether anyone is home.