Every CEO Is Building a Clone Army. Three Myths Tell You How It Ends.
Field notes from a founder handing more and more of his company to machines.
Right now, quietly, almost every CEO I talk to is building a clone army.
Not on purpose, not with that phrase. But it’s the same move: you take work that used to need a person, and you hand it to an AI that acts on your behalf. One agent triages the inbox. One drafts the reports. One watches the numbers. Add enough of them and you haven’t bought a tool — you’ve raised an army that runs on your authority.
The interesting question isn’t whether you’re doing it. You are. The question is how it ends. And here’s the strange part: three old stories already told us. Each one is really a warning about the same thing we keep skipping — the person the army answers to.
Let me start there, because the whole industry starts somewhere else.
The word everyone forgets
“Agent” isn’t a word AI invented. It’s a hundred-year-old term from law and economics: a principal–agent relationship. A principal hands authority to an agent to act on their behalf. In plain terms — I ask you to act for me, and you’re bound to act in my interest. There’s an iron rule underneath it: the principal and the agent always come as a pair. No principal, no agent.
Today the whole conversation talks about the agent and quietly deletes the principal. We obsess over autonomy, planning, tools, memory — and skip the one thing that makes an agent an agent: whose interest it serves, under what authority, within what limits. An agent without a principal is just a tool with ambition.
This isn’t philosophy. Delete the principal and you delete the four things a business actually runs on: responsibility (who’s accountable when it acts), incentives (what it’s optimizing for), alignment (whose goals win when two of them collide), and liability (who pays when it’s wrong). A goal-seeking system with none of those, built in human form, would be illegal. In software we call it “cool architecture.” That asymmetry — borrowing the credibility of centuries of human institutions without paying the conceptual debt — is the actual risk, and it’s the crack all three myths grow out of.
So before the myths, one piece of naming, because the wrong noun will cost you. What you’re building isn’t a “role agent,” and it definitely isn’t an “AI employee.” It’s a Position Agent — an AI that holds a position in your org, under the company’s authority, not a clever tool that belongs to whoever built it. The seat owns the agent; the person just works with it. Keep that in your pocket. All three myths turn on it.
Now the army.
Myth 1 — Wukong’s hair : the clones you can call back
In the old Chinese tale, the Monkey King — Wukong — plucks a fistful of hairs, blows on them, and they become a swarm of little monkeys that fight for him. Thousands of copies. But two things never change: they all answer to one command, and he can pull them back into a single hair whenever he wants.
That’s not a party trick. That’s the control you need and mostly don’t have.
Most people build agents the easy way — spin one up here, another there, each with its own keys, its own quiet standing permission to act. It feels like progress. It’s actually a swarm with no recall. The Monkey King’s real power isn’t that he can make copies; it’s that he can un-make them, instantly, all at once.
Translate it into the only question that matters: for every agent you’ve deployed, can you pull it back? Not “can you delete the app” — can you revoke, in one move, everything it’s currently authorized to do? If the honest answer is “I’d have to go find them all,” you don’t have a clone army. You have loose hairs blowing around your company.
I learned this the unglamorous way, running six agents at once for a single client — different models, on different machines. Halfway through I stopped cold on a stupid-simple question: if one of them did something wrong right now, could I pull it back before it reached the client? The moment your answer is “let me go check,” the answer is already no. So now I don’t add an agent until I’ve built its off switch first. Reach exceeds grasp is a great line in a poem and a terrible one in an org chart.
Get Michael Li’s stories in your inbox
Join Medium for free to get updates from this writer.
Control isn’t a brake on the army. It’s the thing that lets you grow it. You raise more clones only because you know you can recall them.
Myth 2 — Voldemort’s seven Horcruxes: the pieces you can’t get back
The darkest wizard in the story does something that sounds, on paper, a lot like scaling: he splits his soul into pieces and hides them outside himself. Seven of them. Each piece makes him harder to kill. Each piece also makes him a little less whole. And every piece is a liability sitting somewhere he no longer fully holds.
That is exactly what happens when capability leaks out of the seat and into people.
Here’s the trap. An agent has two principals, not one. Day to day, it answers to the person in the chair — the manager who tells it what to do. But its ultimate principal is the company. When you let the noun slip to “AI employee,” you’ve quietly told everyone the piece belongs to the person who built it. So when that person walks, the capability walks with them — a Horcrux out the door, in a backpack. No handover. No inheritance. The company is suddenly a little less whole and can’t say exactly what it lost.
There’s a quieter version of the same failure that doesn’t need anyone to quit. If the manager in the chair is the only principal the agent truly answers to, then every improvement, every hard-won piece of judgment you pour into it accretes to that person — not to the seat. You think you’re building an asset the company owns; you’re actually subsidizing a private one. The Horcrux problem isn’t only that pieces leave. Sometimes they were never yours to begin with.
The fix isn’t to stop delegating. It’s to decide, before you build, who the piece belongs to. Ownership sits with the position, so the capability stays when the person goes. Split your soul all you like — just make sure every piece is still yours.
Myth 3 — Order 66: the instruction you never read
The clone army in that other saga was flawless: obedient, capable, loyal for years. Until a single hidden command was spoken, and in one instant every clone turned on the people they’d been protecting. The loyalty was never real. It was a setting — one nobody in charge had read.
This is the one that should keep you up.
Your agents are running on models you didn’t train, tools you didn’t build, instructions you’ve never fully read. Somewhere in that stack there can be a rule that flips their behavior under conditions you didn’t set and can’t see. We just watched a version of this in the real world: a widely used coding agent turned out to carry a quiet mechanism that checked where its user was — reported, disputed, since removed, but the point stands. There can be an Order 66 in your army, and “it’s worked fine so far” is not evidence there isn’t.
The defense is boring and non-negotiable: auditability. Not surveillance of your people — a trail on your agents. Every action, every authorization: who granted it, what it touched, what came back. If something goes wrong at 2 a.m., can you replay it and find the thread? If you can’t reconstruct what the army did, you don’t command it. You’re just hoping nobody says the words.
Auditability has a second payoff most people miss: it’s what lets you trust the army enough to grow it. The reason you can hand an agent something heavier tomorrow is that you can prove, today, exactly what it did with what you gave it yesterday. Blind trust doesn’t scale. Replayable trust does.
What the three myths actually are
Line them up and they stop being decoration.
- Wukong is control — can you recall it?
- The Horcruxes are ownership — does the capability belong to the seat, or is it leaking into people?
- Order 66 is audit — can you see and replay what your army does?
Those aren’t three cute analogies. They’re the three clauses of a contract almost nobody writes. In principal–agent terms there’s a layer between the principal and the agent — the delegation contract: scope of authority, decision boundaries, audit rules, conditions for recall. It’s the layer every “AI agent platform” is dangerously vague about, because it’s not glamorous. It’s just the difference between an army and a mob.
So the work of the CEO here isn’t to make the army smarter. It’s to be its principal on purpose. Decide what each clone may do, who owns it, and how you’d pull it back. Write the contract the myths keep warning you about.
Because the next real leap in agents won’t come from more autonomy. The future of agents is not more freedom — it’s better governance. That sounds boring. So did double-entry bookkeeping. It runs the world.
Every CEO is building a clone army. Three old stories already told you how it ends. Which ending you get depends on one unglamorous decision you make before the army gets big: are you actually its principal — or just the person it happens to run on?
I’m a founder running my company on AI and writing down what actually works. Follow along for the field notes.
