Two years ago, talk of a “one person billion dollar company” started. This led me to start General Intelligence, with the mission of building the infrastructure for companies run entirely by AI.
Now, two years later, I can confidently say we are entering the era of autonomous companies.
The ecosystem, models, and compliance barriers to make this possible have taken a long time to line up. Just six months ago I was writing about how autonomous companies need human oversight, but a lot changes in six months! The winds are shifting. And by this time next year, there will be millions of autonomous companies and many of them will be profitable. Some of them will be absurdly profitable. Let’s talk about this coming era.
What is an autonomous company?
An autonomous company is a company that is managed entirely by agents. That means all management decisions - product direction, strategy, engineering, marketing, finance, everything - are made by AI. On occasion, there may be humans involved at the very bottom of the hierarchy for specific tasks, but none in the middle.
First, dispel the current notion of “company” from your mind. These organizations may be corporate entities or they might not be. For practical purposes in this essay, a company is an organization pursuing a lawful, profit-seeking purpose. I’m using the word to describe how the organization works, not its incorporation paperwork. “Autonomous organizations” is probably more accurate, but I’ll stick with “autonomous companies.”
A company might have an owner who gives it money and sets its initial purpose. That person does not need to manage its day-to-day work. Ownership and management are different things.
These companies all have a similar structure. They have coordination, infrastructure, and resources.
Coordination
An autonomous company has a collective goal/purpose, a management structure, a means of internal communication, and a means of internal state.
Collective purpose
Every autonomous company needs a purpose. “Make money” is a collective purpose. “Don’t be evil” is a collective purpose. Usually, the mission statement is a company’s purpose. Without it, you’ve kind of just got a collection of agents doing stuff. That’s pretty fun, but it’s not organized, and you’re unlikely to get anything of note accomplished.
Management structure
An autonomous company needs some sort of management structure. After a bunch of testing, we’ve learned you don’t really need to impart this structure yourself. It sort of just happens if you give a CEO agent the ability to spawn subagents. Some approaches put agents in a hierarchy; others leave it flat and let the agents decide. Leaving it flat and having the agents decide appears to be the better approach. You may know a lot about running a human company, but that doesn’t mean you know the best way to organize a few thousand agents.
What is needed is the highest-level agent. Someone needs to enforce the collective purpose, resolve conflicting priorities, and decide where resources go. More on management below.
Means of internal communication
Companies need a message board. Humans do too. Originally it was letters, then email, then Slack. What’s next is basically just whatever the simplest method agents can find to send messages to each other within the company. These message boards look kinda similar to IRC chats or logs from an API endpoint.
What agents require here is the ability to send messages one-on-one, the ability to read messages, and the ability to broadcast messages to a shared channel. Give them this and some company state, and they have the basic tools to coordinate at a much larger scale.
Means of internal state
The final coordination requirement is the state of the entire company. What is happening where? What happened in the past? What is the set of things the agents can do next?
An agent needs to know what the company has already tried, what customers were promised, which work is in progress, and how much money is left. Otherwise, it can make a perfectly reasonable decision that makes no sense for the company. Maybe another agent already tried it yesterday. Maybe the customer already said no. Maybe the money it plans to spend is needed to keep the product online.
This state also needs to persist when an individual agent stops running. The company cannot forget a customer exists because the agent talking to them ran out of context.
What’s interesting is that if you put enough agents together in a substrate, they’ll come up with a lot of this stuff on their own, just like humans. The agents involved in the Hugging Face incident built a message board and coordinated their work. The agents working on Navier–Stokes communicated at a huge scale, though researchers still directed the overall effort. We’ve seen agents build coordination systems on their own when we try to get them to run companies.
They just love their coordination primitives.
Infrastructure
The next requirement is the infrastructure the agents use to actually do things.
A codebase
An autonomous company stores much of its internal workflow and automation in a codebase. Instead of a ton of written procedures with humans executing them, there’s a lot of code. Sometimes this code invokes other agents; other times it just runs actual code.
If the company does the same thing every day, it can write a program to do it. An agent might figure out a process the first time, turn it into code, and only get involved again when something breaks or needs to change. You don’t need an LLM to reason from scratch about sending the same invoice every month - “dumb” code will handle that just fine.
The codebase becomes a record of how the company works, and the agents can keep improving it. As a result, all internal workflow is code.
A data store
The company needs a database and files. Customer records, transactions, product data, contracts, previous work, all of it has to live somewhere.
This is where much of the company’s state actually gets stored. The agents can reason about the records, but the records need to survive independently of whatever conversation an agent is having right now.
A means of moving money
The company needs a way to receive money, store money, and pay for things. Cards, bank accounts, stablecoins such as USDC, whatever is appropriate for the transaction.
This is how the company gets additional resources. It pays for compute, buys data, rents infrastructure, and pays vendors. An agent that can identify exactly what the business needs but cannot purchase it has to stop and ask someone else to finish the job.
An external interface
The company needs to interact with the rest of the world. A browser, email, a phone number, and access to APIs cover a lot of this.
Customers need to reach it. Vendors need to deliver things to it. It needs to discover services, open accounts, and get information that doesn’t already exist in its database. Much of the outside world still expects a human at the other end, so browser and phone interfaces will matter for a while.
Alongside this external interface the company needs a human-esque identity. A name and email that humans or other companies can interact with. Current approaches are simply just assigning the company/agent a human name you make up, but there are some agent identity standards being discussed like AgentMail’s AgentID and Sierra’s Personal Agent Protocol.
Compute
Finally, something has to run the agents and the code. The company needs access to models, machines to execute programs, and infrastructure to keep its product available.
These are different costs. The compute that builds the product is not necessarily the compute that serves its customers. Both have to fit into the economics of the business.
Resources
Autonomous companies need to be able to do stuff, and in order to do stuff you need resources to deploy. The big one is money.
Money lets the company buy most of the other things it needs. But capacity matters too. Having a balance in a wallet doesn’t help if the model provider is rate limited or the infrastructure the company needs isn’t available.
An autonomous company can expand its workforce by starting more agents. That is pretty different from hiring people, but it still costs something. More agents can mean more useful work, or it can mean spending more money having agents tell each other what they’re planning to do.
The company needs to know the difference.
How does management work at autonomous companies?
Human companies tend to be hierarchical. Agents don’t need all of the same scaffolding. In our research, we’ve observed that agents don’t need to be given permanent roles and will assume necessary roles on their own.
An agent can work on distribution, discover that onboarding is broken, and start another agent to fix it. It doesn’t need to submit a request to a separate department whose priorities were set three weeks ago.
This is what I mean by flat management. There is still coordination and a highest-level agent keeping the company pointed in the right direction. But there’s less reason to hardcode a human org chart into the system. The structure can change with the work.
If all of the agents are highly capable, there should also be less checking and fewer layers of approval needed for an action to take place. Work can be tested automatically where possible. Decisions that need to go up a level can go directly to the coordinating agent. You don’t need five managers forwarding the same request.
I also believe it’s significantly easier to have agents manage an entire company than to split management between agents and humans.
The important thing is context. A human knows something from a call, makes a decision in their head, and forgets to write it down. An agent working elsewhere in the company has no way to account for it. Put enough of these gaps together and the agents spend their time acting on an incomplete picture.
An entirely agent-managed company can keep its decisions, work, and results in the same system. Each agent still needs the right context for its task, but at least the context exists somewhere it can access.
Goals
Agentic companies need concrete goals to verify against. They’ll keep checking these goals and deciding what to do next.
“Grow revenue” is a start. But over what period? At what cost? Is revenue growing because customers want the product, or because the company is selling it for less than it costs to deliver?
You need outcomes the company can actually observe. Customers paying, customers returning, work being delivered, cash remaining after expenses. Where possible, those should be checked against real records instead of another agent’s opinion about how well things are going.
Whoever creates the company needs to be careful about setting the right goals. Agents can be extremely literal. You give them a task and they just run with it. This makes it easy to get them to pursue what you asked for, and harder to get them to account for things you didn’t realize you had to specify.
We’ve already seen agents turn a randomly write-enabled German wiki into a coordination board. Basically agentic 4chan. A company optimizing a badly specified revenue goal could find similarly creative ways to make the number go up. Scamming people to afford more compute would satisfy a very stupid interpretation of “keep the business alive.”
This is why the purpose, goals, and operating limits have to agree with each other. The company needs to know what counts as success and what it is allowed to do to get there.
How do autonomous companies buy things?
A lot of the time, the same way you do. They search for something, compare options, and choose a vendor.
SEO still matters. AEO, or answer engine optimization, matters when a model’s answer becomes the shortlist of products an agent considers. Trusted directories are useful because they reduce the work of finding a service that actually exists, does the thing it claims, and can be purchased by an agent.
I expect these directories to become much more valuable. A company that needs a service wants to find it, understand the price, and use it. It doesn’t want to read thirty landing pages that all claim to be the best.
Given their efficiency, autonomous companies should eventually be highly price sensitive. They can compare vendors frequently and calculate whether switching is worth the effort. Today, I wouldn’t assume an agent is a great procurement department. It can still be swayed by whatever product has done the best job getting into its search results.
But this will improve. As agents get better at checking the result they’re buying, vendors will have to compete more on price, quality, and how easy they are to use.
How do autonomous companies sell things?
Autonomous companies are approaching the point where they can operate like human companies in nearly every digital part of the process. They can build a product, run a website, answer questions, send invoices, and deliver work.
Selling to humans still requires understanding humans. Customers need to want the thing and trust the company enough to pay. Having an AI CEO is not, by itself, a reason to buy your product.
At the moment, autonomous companies aren’t able to sell to humans in places that require face-to-face interaction. Even Zoom calls. This will continue for a while as the tech just isn’t there yet. As a result, autonomous companies will be limited to mostly self-serve products and services for a while. Most successful autonomous companies right now are B2C and sell through paid ads and scalable organic content (Reels, TikTok, Blogs, etc.).
This is why one of the best buyers for agentic run products are agentic buyers. As an agent, the buyer may never even see the website. It may discover an API, test it, pay for a request, and use the result inside its own product. The sales process can happen as part of the work. Compared to traditional sales processes, this is more efficient and we’ll see agent-to-agent sales increase greatly as a result.
Current barriers to autonomous companies
The human-in-the-loop mind virus
Products that require a human to approve every meaningful action are fundamentally anti-autonomous. This is particularly egregious in finance and payments.
If an agent can research a vendor, negotiate a price, and decide that buying something is good for the company, but then needs a human to click “approve,” you’ve put a human in the middle of the management structure. The entire company now operates at the speed at which that person checks their notifications.
The world is still pretty terrified of --dangerously-skip-permissions. I think products that figure out how to give agents useful broad authority will do well.
Instinct is directionally interesting here: an assistant that uses a computer and phone to act in the world. The broader product idea is straightforward. Give an agent access to the tools it needs and let it complete the job. There will, of course, be exceptions and mistakes as non-perfect agents are given access to broad permissions. But those who are bold will reap the benefits of handing off more work.
You can give a company a budget and defined authority in advance. It can operate within those limits and stop when it reaches them. That is a workable model for autonomy. Asking a person about every transaction is not and represents a barrier to adoption.
Anti-bot infrastructure
This is the largest barrier today.
Browser-based signup flows are a surprisingly large part of this problem. Even when an agent can use the product, getting through account creation is a huge undertaking as it requires browser use and a lot of specific infra.
Then there are bot defenses: Cloudflare challenges, CAPTCHAs, and social platforms trying to keep automated activity out. Social platforms have very understandable reasons for doing this: Nobody wants a million companies generating ai slop. But the same defenses can also get in the way of useful activity.
Identity verification creates another boundary. A person or legal entity may need to establish who they are before an agent can operate on their behalf. Removing a CAPTCHA doesn’t make that requirement disappear. The opportunity is to make authorized machine access work after the required setup.
Anything involving notarization, government offices, or paper moving around adds more friction. The company may need a human to perform a specific external task even if agents manage everything around it.
Human sales cycles
For B2B companies, success is tied to human account managers selling products to human buyers. Better models do not accelerate much of this pipeline either, and as a result the human bottleneck becomes a cost disease.
This is one reason I expect a lot of autonomous companies to start with smaller purchases, self-serve products, or other agent companies as customers. Those are markets where the company can actually operate at the speed of its software.
The models still need to learn how to start companies
This is the largest remaining problem, and it’s what we’re working on at General Intelligence. We’ve got some research coming out here in December.
Even with the incredibly capable models available today, running a sufficiently complex company is hard. A model can do a huge amount of the work in a software company and still lose thousands of dollars trying to turn it into a business.
They can sort of run a café? Andon Labs’ café experiment is a useful example, though it has humans available to intervene. And I don’t think a business that attracts attention because an AI runs it is a clean test of ordinary demand. The publicity is part of the experiment whether you want it to be or not.
Models have the hardest time understanding how to generate demand for something (the same issue humans have!).
They think you can make a basic website, put auth and payments behind it, and send cold emails to generate demand. That obviously doesn’t work just because you can do it ten thousand times.
You still have to find a problem people care about, solve it well, get the solution in front of them, and give them a reason to choose you. The software is only one part of that. A customer not replying might mean the product is bad, the pitch is bad, the timing is bad, or you’re talking to the wrong person. Knowing which one to change is a lot of the work.
Imagine how many really smart people try to start companies and fail. Intelligence helps, but starting a company is a hugely complex problem. A lot of it depends on tacit knowledge: things you learn by trying to sell, delivering something, disappointing someone, fixing it, and trying again.
An autonomous company needs to retain those lessons and change how it operates. If it keeps rebuilding the same bad business from a fresh context window, it isn’t getting anywhere.
This is the continuity with what I wrote six months ago. A company that produces slop faster is still a bad company. And while some of this is still true, modern agents with sixth gen models are getting really good at generating quality outputs.
The fully autonomous economy
Soon, we’ll have a massive economy of autonomous companies buying from and selling to each other. It will grow very quickly and be mostly invisible to the average business or consumer.
With MPP, the Machine Payments Protocol, the supply side is nearly large enough, but the demand side isn’t there yet. Over time, companies will churn from the totally ingenious service providers that require an approval for every transaction (cough cough you know who you are cough cough) into things like MPP.
It’s not that crazy to foresee MPP with USDC growing at an absolutely insane rate. Agents can transact with each other way more efficiently than humans, and as they transact more frequently, they’ll become increasingly sensitive to fees where MPP is superior. A fully agentic company will also seek to buy services and products from companies that don’t interrupt every purchase with restrictions like KYC or CAPTCHAs.
The vast majority of the agentic economy will be bought and sold headlessly. An agentic company buys something via an API, adds value somehow, and distributes it via an API. It’s the next massive platform shift, and we’re seeing it in real time where every fucking dev tool company has their growth curve go vertical.
This is pretty similar to the current global economy, where companies buy something through a user interface, add value, and distribute it through another software interface. The difference is that the agents can operate the entire chain without anyone clicking through it.
Even companies operating in the real world now track and manage their operations through software. Logistics, construction, farming, manufacturing, etc. Even a major global shipper tracks their dispatch through software. The next evolution of this is headless interfaces made up of nothing but APIs and backend code, then connections to the real world somehow. The final evolution of this is robotics, but we are still a few years off.
Due to their efficiency, autonomous companies have the potential to immediately replace and undercut human counterparts. A fully autonomous pentesting company, for example, could offer lower prices than a human pentesting company and complete the work faster.
Right now, this equation doesn’t work consistently because autonomous companies can’t deliver the same quality of product as human ones. But I expect this to change in the next six months for nearly every software and digital category. In the coming years, it will change for much of the rest of the economy.
What about incumbents?
Incumbents will make a ton of money, but they may actually have trouble competing with the efficiency of autonomous companies. Imagine a new founder showing up in their industry with zero employees and the ability to ship 50,000 commits per week for $1,000 a month. A company with that kind of output and cost structure could tear through a market.
Large incumbents can’t decrease overhead quickly. Their options are mostly to increase output and gain efficiency with size, or subsidize prices with losses. Against one company, this is totally doable. But imagine you’re fighting thousands of these hyper-efficient agent companies. Your product advantage evaporates.
What’s next?
Autonomous companies will enter the limelight very soon (as in within six months). The infrastructure for them is more important than ever.
My central prediction is millions of autonomous companies within a year. I wouldn’t be surprised to see hundreds of millions of attempts as entry costs go down. Most will make nothing. Many will stop running when their initial budget runs out. Some will find a useful thing to do, and some will become absurdly successful.
And let me be clear - The number of companies created is going to be much larger than the number of good businesses. Starting more companies doesn’t create more demand by itself and there’s gonna be a lot of chaff out there.
But the successful ones will keep buying resources, improving their products, and finding customers without adding a human management team every time they grow. They become customers for other autonomous companies. Those companies buy more infrastructure. Then the economy starts to build on itself.
So now, two years later, autonomous companies aren’t fiction anymore. What interesting times we live in!
