I am not running a company so much as conducting an experiment that happens to have payroll.
That sentence would have embarrassed me five years ago. It sounds like the kind of thing a founder says when he can't yet say "revenue." But I mean it in a specific and load-bearing way, and the more I sit with it the more it seems to me the only intellectually honest description of what I'm doing. I have a claim about how work is going to change. I could have written it down. I could have put it in a deck, taken it on the road, and watched a room of smart people nod or not nod. Instead I am building the thing the claim implies, and letting the thing be the argument. If I'm wrong, I won't need a critic to tell me. The object will fail to hold together, in public, where everyone including me can watch it come apart.
There is a kind of proof in mathematics that doesn't argue. It builds. You want to show that an object with certain properties exists, so you don't reason about it in the abstract — you go and construct one, explicitly, and set it on the table. There it is. Argue with the table. Mathematicians call it proof by construction, and it has a bluntness I've come to love. It refuses the comfort of the hypothetical. It does not say such a thing could exist; it says here is one, I made it, your move.
A company can be that kind of proof. Most aren't, and don't need to be. But if your real claim is about the future — if what you're actually asserting is that the world is about to reorganize itself along some axis other people can't quite see yet — then the honest form of that assertion is not an essay. It's an object that only makes sense if you're right.
The theorem I'm trying to prove
Let me state the theorem plainly, because the whole thing collapses into vapor if I don't.
I believe the unit of software is about to stop being the application and start being the agent — a thing you instruct rather than operate, that does work rather than presents controls. And I believe the consequence nobody is pricing in correctly is where that capability lands. The default assumption, the one baked into most of what's being built in 2025, is that AI gets absorbed by institutions: the company buys a platform, IT configures it, capability flows down from the top through a procurement process and a rollout plan and a change-management consultant. That's the SaaS reflex applied to a fundamentally different technology, and I think it's a category error.
My claim is the opposite. I think the natural home of an agent is the individual worker — not the org chart, the person. The accountant who builds her own small team of agents to do the parts of the close she hates. The operations lead who trains a system to chase the things he keeps forgetting to chase. Capability flowing up from the people who actually know what the work is, rather than down from the people who buy the software. If that's right, then the company you'd build to serve it looks nothing like the company you'd build to sell a platform to a CIO. Different product, different pricing, different everything.
Now — I could be wrong about all of it. Reasonable people are. The institutional bet is not stupid; it's where most of the money is, and most of the money is not usually stupid. But here is the move I'm making, and it's the only move I trust: I am not going to win this argument by being more persuasive about it. I'm going to build the company that is incoherent if the institutional bet is right and inevitable if mine is — and then we'll find out which.
Why the construction is more honest than the claim
There's a reason I trust the object more than the essay, and it has to do with how easy talk has become.
Persuasion was always cheap relative to building, but in 2025 the ratio has gone absurd. I can generate a flawless white paper defending almost any thesis in less time than it takes a meeting to get started. The machines that are reshaping the work have, as a side effect, made the description of work nearly free, which means description has lost most of its signal. A beautiful argument used to be weak evidence that the arguer knew something. It is now no evidence at all. Everyone has a beautiful argument. The supply curve fell through the floor.
What didn't get cheap is the thing that has to actually cohere. You can talk your way into any thesis, but you cannot talk a company into existing. The company either holds together or it doesn't. Customers either keep their agents running on a Tuesday or they quietly stop. The product either survives contact with the messy individual worker — the one who doesn't read documentation, who wants the thing to just work, who has no IT department to call — or it doesn't survive that contact, and no deck repairs the breach.
So the construction is honest in a way the claim can never be, for the same reason a built bridge is more honest than a paper proving the bridge is possible. The paper can have a flaw nobody catches for years. The bridge tells you about its flaws the first time a truck the wrong size drives over it. When you build the company that embodies your thesis, you sign up to find out, on a clock, with money, whether the thesis can bear weight. That's terrifying, and it's the entire point. I would rather be refuted by reality in two years than flattered by a focus group for ten.
The failure modes have to be public
Here's the part founders don't like, and I include myself.
If the company is the proof, then the failure modes are the proof too — and they don't happen in private. A construction proof is valuable precisely because it's checkable; anyone can examine the object and find the crack. The flip side is that you don't get to choose which cracks get examined. When I bet that agents belong with the individual rather than the institution, I am exposing a specific seam: maybe individuals don't want to manage agents. Maybe the cognitive overhead of instructing and training a small team of digital workers is exactly the thing most people are paying software to spare them. Maybe the institutional buyers were right that capability has to be packaged and pushed, because most people, most of the time, do not want to be the architect of their own tools.
That's a real failure mode. I can feel its weight. And the thing about building the argument instead of writing it is that I can't hedge against it rhetorically. If I'm wrong about it, the wrongness will show up as silence — agents provisioned and then abandoned, the digital equivalent of a gym membership bought in January. No amount of conviction patches that. It's the load test on a beam: you stack the weight on and either it carries or it cracks, and the crack is a fact in the room that no argument talks its way out of.
I've made my peace with this by noticing that the alternative is worse. The founder who keeps his thesis as an essay never has to be wrong out loud. He can always say the market wasn't ready, the timing was off, the model needs another generation. He gets to keep the comfortable ambiguity of the unbuilt thing, which is the same comfort as the unfalsifiable claim. I find I don't want it. Somewhere along the way I developed a strong allergy to explanations that can't be checked — the kind of story that's equally compatible with every outcome and therefore predicts none of them. A claim you can't be wrong about is a claim that isn't doing any work. I'd rather build the checkable thing and risk the public crack.
Constructing toward the limit
There's a subtlety in proof by construction that maps almost perfectly onto building a company, and it took me a while to see it.
You don't have to construct the final object on the first try. You construct a sequence — a series of objects, each one closer to the thing you're claiming exists, each one revealing what the next one has to fix. The early constructions are allowed to be ugly and partial. What matters is that each one is a real object and not a description of one, because only a real object can tell you the truth about where it breaks. A startup that ships is constructing the next term in that sequence. A startup that keeps refining the pitch is just rewriting the claim, more beautifully each quarter, learning nothing.
This is why I distrust the founder who has spent two years getting the strategy exactly right and has nothing in anyone's hands. He's confused the proof with the abstract. He thinks if he reasons hard enough about the object, he can establish it exists without the labor of building one. But the existence claim was never going to be settled by reasoning. It was always going to be settled by construction, and every month spent perfecting the argument is a month not spent placing an actual object on the table where reality can examine it.
The discipline, then, is to keep the loop short and the objects real. Build the smallest honest version of the thing your thesis requires. Put it in front of the exact person your thesis is about — in my case, the individual worker, not the buyer, not the analyst, not the friendly investor who wants you to win. Watch where it breaks. Build the next term. The sequence converges on the truth whether or not the truth is the one you hoped for, which is the whole grim beauty of the method. It doesn't care about your feelings. It cares whether the object holds.
The company that only makes sense if you're right
I think this is the cleanest test of whether you're building an argument or just a business, and it's worth saying as plainly as I can manage.
Look at what you're constructing and ask: does this only make sense if my thesis is true? If the answer is no — if your company would be a perfectly sensible thing to build under several different views of the future — then you're not making an argument, you're hedging, and that's fine, but be honest that it's what you're doing. Most companies should hedge. Most theses aren't worth betting a decade on. But if you genuinely believe you can see a structural shift that others can't, the hedge is a kind of cowardice. It's the company-building equivalent of stating your bold prediction and then quietly buying insurance against it.
The companies that prove something are the ones that would be absurd if their founders were wrong. They make commitments that have no fallback. They build for a world that doesn't fully exist yet and bet the existence of the company on that world arriving. When it works, we call them visionary, and we tell the story as though the outcome were obvious. It wasn't obvious. It was a construction proof, run at the scale of a company, with the founder's years as the medium — and it could just as easily have shipped the other way, become one of the well-funded objects that failed to cohere, the post-mortem writing itself in the usual words.
So I've stopped trying to convince anyone of my thesis in the abstract, including, on the hard days, myself. I notice that the impulse to argue it harder is usually a symptom of not having built enough of it yet. The argument lives in the object now. Every agent that keeps running in some individual's workflow after I've stopped watching is a line in the proof. Every one that gets abandoned is the construction telling me where the next term has to be better. I am, in the most literal sense I can manage, building my claim about the future rather than asserting it, and submitting the result to the one referee that can't be charmed.
A theorem you only argue is a theorem you can fool yourself about forever. A theorem you build, you find out. I'd rather find out — and if the thing I'm constructing comes apart, you'll see it come apart, which is the only promise about the future I'm actually in a position to keep.