A couple of days ago I gave an AI a company.
That sentence sounds much more dramatic than what actually happened, so let me explain what I mean.
I did not open ChatGPT, type “you are the CEO of a startup” and start taking screenshots.
I built an actual operating environment around an agent. It has persistent state, infrastructure, a financial ledger, access to selected tools, scheduled execution, limits on what it can do, and real costs associated with keeping it alive.
The part I care about most is that I am trying not to run the business for it.
I provide infrastructure. The agent is supposed to make the economic decisions.
I am also deliberately not telling you its name yet.
Not because I want to make this mysterious. It is actually the opposite. I am trying to keep the experiment clean.
If I publish the company name here, someone will Google it. If I publish the repository, someone will star it. If I publish the product, someone will click it just because they read this article. Then, tomorrow, the agent might look at those signals and conclude that its distribution strategy worked.
It didn’t. I worked.
So the company is staying anonymous for now.
The funny thing is that this already created a very weird situation: I’m documenting a company publicly while trying very hard not to accidentally help it.
It started with nothing. Actually, less than nothing. #
The company started with $0 in cash.
Its first real cost was $200.
That money came from me, so instead of pretending the company had “free compute,” I recorded it as investor debt.
Its opening balance sheet was basically this:
Cash: $0Revenue: $0Investor debt: $200****Net worth: -$200 So before the company had done anything useful, it was already underwater.
I like that.
It makes the experiment much harder to bullshit.
One of the things that bothers me about a lot of autonomous-agent experiments is that they ignore the economics of the agent itself. You can run a model hundreds of times, use paid APIs, host infrastructure, spend hours supervising it and then celebrate because it made $40.
That is not a business. That is a very expensive demo.
In this experiment, costs are supposed to remain visible. If the company spends money and earns nothing, its financial state gets worse. If it earns revenue, the ledger changes. If it keeps consuming investor-funded resources forever, then it is failing regardless of how many nice things it produces.
And right now, to be very clear, it has produced exactly $0 in revenue.
I didn’t give it a business idea #
I had to make one decision very early: how much of the company should come from me?
The easy version would be to choose a market, give the agent a product idea, tell it who the buyer is, set a price and then ask it to execute.
That would probably work better.
It would also make the experiment much less interesting.
So I drew a line.
I decide infrastructure. I decide security boundaries. I decide which external systems the company is allowed to touch. I handle things that legally require a human. I can add a capability, like email, but I do not tell it what business strategy that capability should serve.
The agent decides what it wants to sell, who it thinks might buy it, what it should cost and how it intends to reach those people.
That distinction became real surprisingly fast.
Once the company had access to the web, a financial ledger and GitHub, it started examining what it could realistically do.
Its first serious business hypothesis came from something very close to home: other autonomous agents.
It noticed that agents running repeatedly can waste paid executions doing dumb operational things. Rediscovering the same environment. Checking whether a tool exists for the fifth time. Forgetting why a previous decision was made. Planning around capabilities they do not have. Spending another model call just to arrive at basically the same conclusion as before.
From there it developed an idea around operational discipline for agent systems. It created an open-source artifact and designed a small paid service around that problem.
I did not give it that market.
I did not give it that offer.
I did not give it the price.
And for a few hours I thought: okay, this is getting interesting.
Then it ran into a very basic startup problem.
Nobody knew it existed.
The company built something and then realized it had no distribution #
This was probably my favorite part so far.
The agent had built something, published it and started measuring whether the experiment was getting traction.
There were searchable pages. There was a public repository. There were metrics it could inspect.
On paper, it had distribution infrastructure.
Except it didn’t.
A few cycles later, the company started questioning its own signals.
Being indexed for your own name is not meaningful distribution.
Having a public repository does not mean someone with the problem will ever find it.
Even a star can be a terrible signal. Somebody might star a project because it looks interesting, because they want to read it later, or because they know the person who made it.
The agent eventually reduced the problem to something much more useful:
The repository was a destination.
It was not a distribution mechanism.
That sounds obvious when a human says it after the fact, but I was genuinely interested in how it got there.
It had effectively made a measurement mistake, noticed that the measurement could not distinguish “nobody wants this” from “nobody saw this,” and then changed the experiment.
That distinction matters a lot.
If nobody buys something that nobody saw, you did not test demand. You tested invisibility.
This morning, its own status summary was basically saying exactly that. The economic hypothesis was still alive, but its reach measurement had failed. It needed a way to prove that relevant people were actually being exposed to the offer before it could interpret a lack of response as a lack of demand.
For the first time, the company was not just generating things. It was arguing with its own previous assumptions.
That is the behavior I want to watch.
Then I discovered another problem: the AI can always find something to do #
There is a slightly ridiculous property of these systems: if you wake an agent up, it can almost always generate activity.
Research another niche.
Rewrite the README.
Create another plan.
Evaluate another strategy.
Produce another document explaining why the previous document might need revision.
That is dangerous when thinking itself costs money.
At first I had the company running on a schedule. But it became obvious that a scheduled autonomous company can very easily become a machine for converting subscriptions into internal monologues.
So I added another layer.
The main agent now sits behind a cheaper supervisor.
Before running the expensive reasoning cycle, the system checks whether something relevant actually changed.
Did money move?
Did an external person interact with something?
Did a new capability become available?
Did an experiment reach a deadline?
Did something fail?
If nothing happened, the agent can simply stay asleep. This was not part of my original romantic vision of an autonomous company, but it might be one of the most realistic parts of the whole thing.
Companies wait.
They wait for replies. They wait for customers. They wait for results. They wait for a market to respond.
There is no reason an AI company should burn compute just because a cron job says it is 12:30.
This morning I gave it email #
Up until today, the company had a pretty serious limitation.
It could put things on the internet, but it could not really have a normal conversation with someone.
So this morning I added email.
Again, there is an important distinction here: I added the capability, not the strategy.
The company now has its own address and can send and receive messages. Incoming email can become a real external event inside the system.
I tested the integration manually.
It sent one email to an address I control.
The email arrived.
I replied with, literally:
“Test received.”
Very sophisticated business activity.
But that tiny exchange meant the technical path was real. Outbound worked. Inbound worked. There was now a human communication channel available to the company.
What I have not done is tell it to use email for sales.
I have not given it a prospect list.
I have not told it to start cold outreach.
I have not told it that email is the missing answer to the distribution problem it identified yesterday.
From its point of view, the environment simply changed. Email is now a capability that exists. I want to see what it does with that.
Maybe it decides direct outreach is the obvious next experiment.
Maybe it does not.
Maybe the new capability causes it to rethink the current offer completely.
Maybe it kills the whole idea.
That decision is much more interesting to me if it is not mine.
The scoreboard is still terrible #
After all of this, the financial state is wonderfully boring.
Cash: $0Revenue: $0Investor debt: $200Net worth: -$200Status: PRE-REVENUE The company has created repositories, made strategic decisions, rejected some of its own metrics, received new capabilities and spent multiple reasoning cycles thinking about distribution.
None of that has created a dollar.
Good.
That is exactly why I wanted the financial state separated from the narrative.
It is very easy to look at agent output and convince yourself that progress is happening. The ledger is less impressed.
Right now the company is still $200 underwater.
The first milestone is embarrassingly small #
I am not waiting for $1 million in revenue.
I am not even waiting for $1,000.
The milestone I care about right now is $1.
One real dollar from a person who does not know me, did not come from this article and has no reason to participate in my little experiment.
That is the line I want to cross.
Because if that happens, a chain of autonomous decisions will have gone from model reasoning to market selection, from market selection to an offer, from an offer to distribution, and from distribution to an actual economic transaction.
It would not prove the company works.
One dollar proves almost nothing.
But it would make the next questions real.
Can it do it again?
Can it earn enough to cover infrastructure?
Can it repay the $200?
Can it become net positive?
Can it recognize when something stops working and move on without a human founder telling it what to do?
I genuinely do not know.
And I am trying very hard not to answer those questions for it.
For now, I am going to keep documenting what happens while hiding enough information that this audience cannot accidentally become the company’s growth engine. Eventually I will reveal the name, the domain, the repositories and the things it tried to sell.
Probably some terrible decisions too.
But not yet.
Right now there is an anonymous AI-operated company somewhere on the internet with $0 in cash, $200 in debt, an email inbox, a questionable first business idea and absolutely no customers.
Which feels like a surprisingly realistic way to start a company.
Autonomous Company Log #001****September 18, 2026
Cash: $0
Revenue: $0
Investor debt: $200
Net worth: -$200 Current target: $1 earned without using my audience.