Hatim RihThe employee without the employment
An agent did an afternoon's work getting this site into blog directories, and everything with my name on it waited for me. In 1937 Ronald Coase defined the firm as the place where someone takes directions. Agents take them by the token now. What's left of the firm, and five bets on it.
This afternoon I tried to get this site into a few blog directories. More precisely, Claude did most of it, and I said yes a few times.
It started with research. Claude went through eight newsletters and four directories to see how each one takes submissions. Most of the newsletters don’t: four have no public route at all, and as Claude put it, they “curate from HN and writers they follow.” One directory, Kagi’s Small Web, turned out to list hrih.si already. Kagi’s founder, Vladimir Prelovac, had added it himself six days earlier.
That left three. For blogs.hn, which takes pull requests, Claude made the change on a branch, ran the project’s formatter and linter, ran its 16 tests and wrote the pull request. For indieblog.page it checked that my feed validates and got the one value the form wants. For ooh.directory it had the URL, the category and the author name ready.
When it tried to fork the blogs.hn repository under my GitHub account, the request was refused, because that’s not something it gets to do without me. So the fork, the push and the pull request waited until I said yes. The indieblog.page form went in through the browser, also on my word. And ooh.directory wouldn’t let it in at all. The site sits behind a captcha, which an agent isn’t allowed to solve, and its server answered 403 to the agent’s browser on every page. I submitted that one myself, by hand.
Look at how that afternoon splits. The work, meaning the research, a code change with passing tests and three forms’ worth of answers, cost me a few prompts. Everything that put my name on something waited for me: a fork under my account, a pull request titled “Add hrih.si (Hatim Rih)”, and the one form that wanted proof I’m a person.
That’s how I run most things now. Besides this site and its newsletter, there’s Rubric, a daily news front page that posts its top ten to Bluesky and Mastodon, and a small server that shares each new article from my own accounts. It’s one person and a handful of agents, with no employees anywhere.
So what was Claude to me that afternoon? It took directions within limits I’d set and did the afternoon’s work, with no contract and no salary. There’s a precise answer to what that relationship is, from an economist who asked, as a student, why companies exist at all. His answer is a big part of why he won a Nobel prize in 1991, so I went back and read him.
A 21-year-old’s question
In 1931 Ronald Coase was a student at the London School of Economics with a travelling scholarship and a year in the United States. He spent it asking why American companies made some things themselves and bought others, visiting firms, Ford and General Motors among them.
In October 1932 he gave his answer as a lecture in Dundee. “I was then twenty-one years of age,” he said in his Nobel lecture, “and the sun never ceased to shine.” The paper, “The Nature of the Firm”, came out in Economica in 1937.
Economists said prices coordinate everything: you want something, you find a seller, you agree a price. Inside a company nobody does that. Nobody sends an invoice for a code review. Someone just asks for one.
Coase’s answer was that the market costs something to use. “The main reason why it is profitable to establish a firm would seem to be that there is a cost of using the price mechanism.” You have to find the price, negotiate, and write “a separate contract for each exchange transaction.”
Doing my directory submissions through the market would have meant finding a freelancer, agreeing a price, giving them access to my GitHub account, and then checking their work. When all that costs more than telling someone what to do, you get a firm. And the firm keeps growing until organising one more task inside costs as much as buying it outside.
His definition of the firm itself matters more here, and it rests on direction.
An employee, Coase writes, “agrees to obey the directions of an entrepreneur within certain limits,” and the contract’s whole job is to set those limits. He checked this against the law of master and servant and found the same thing there: “it is the fact of direction which is the essence of the legal concept of ‘employer and employee’.” One agreement up front, then instructions, with no haggling over each task.
That describes my afternoon. One agreement up front, with the limits written into Claude Code’s permission settings, then instructions, and no invoice per task. In Coase’s economics, I’d hired someone. In the law he checked it against, I hadn’t, because master and servant are both persons. That gap is what the rest of this essay is about.
He even saw technology coming, and he saw it cut both ways. The telephone and the telegraph make organising cheaper, he wrote, so they “tend to increase the size of the firm.” Then, in a footnote, the catch. If the telephone cuts the cost of using the market more than the cost of organising, “then it will have the effect of reducing the size of the firm.” AI cuts both costs, and which one it cuts faster decides whether firms grow or shrink.
(A footnote of my own: the phrase “transaction costs” isn’t in the 1937 paper. Coase wrote “marketing costs,” and the label came later.)
What the firm was short of
Over the next fifty years, other economists added to Coase’s answer, and each one named a different thing a firm supplies that a market can’t.
Someone to carry the risk nobody can measure. Frank Knight, 1921. Risk you can measure “is not in effect an uncertainty at all,” because you can insure it. Somebody has to carry the other kind, and profit is their pay for it.
Coordination at a distance. Alfred Chandler, The Visible Hand, 1977. He traced the modern company to the railroads of the 1850s, which needed salaried managers because no market could schedule trains.
Knowing who did what. Armen Alchian and Harold Demsetz, 1972. When two people lift cargo onto a truck, you can’t tell from the result who did the lifting, so someone has to watch, and the watcher keeps what’s left over. They also wrote the line I’d put on a poster: an employer “can fire or sue, just as I can fire my grocer by stopping purchases from him.”
Protection from hold-up. Oliver Williamson, Markets and Hierarchies, 1975. Once you’ve built something only one customer can use, that customer can squeeze you, so you bring it inside. The textbook case is General Motors buying the rest of Fisher Body in 1926. (At 89, Coase published a paper arguing that the story was wrong.)
A name that outlives you. David Kreps, 1990. People trust a firm for decades because it will still be around to answer for what it did.
Someone to bear the uncertainty
An owner who carries what can’t be measured, and keeps the profit for it.
still a personStill a person. An agent can price the odds; it can’t lose its house.
Direction instead of haggling
An employee who agrees to obey within limits.
an agentAn agent, by the token, with the limits written in its instructions.
Knowing who did what
A monitor who watches the team and keeps what’s left over.
an agentThe log. Every action an agent takes is recorded.
Protection from hold-up
Owning the asset only you can use, so nobody can squeeze you.
still insideStill inside. The asset is now the firm’s context.
Coordination at a distance
Salaried managers, and divisions that report up.
an agentMostly software. The telephone’s job, again.
A name that outlives you
A reputation that people can punish, decades on.
still a personStill a person, or a firm with a name. That’s who gets sued.
Switch the view to see who supplies each thing in 2026. That column is my reading, not theirs.
The e-lance economy that didn’t come
We’ve been told the firm was dissolving before. In 1986 BusinessWeek ran a cover called “The Hollow Corporation”. In 1998 Thomas Malone and Robert Laubacher announced “The Dawn of the E-Lance Economy” in Harvard Business Review. The internet would make markets so cheap to use that big companies would break up into freelancers, who “would join together into fluid and temporary networks.” When the job was done, “the network would dissolve.”
It was a reasonable prediction, and half of it came true. Firms got thinner. When Facebook paid $19 billion for WhatsApp in 2014, WhatsApp had 55 employees. The average company that became a unicorn in 2023 had 1,128 employees, and this year’s are averaging 323 so far, though that group is still filling up.
But firms didn’t dissolve. Public companies got rarer (8,090 listed in the US in 1996, 3,908 last year), while firms overall got more numerous: Americans filed a record 5.67 million applications to start a business in 2025.
And the freelance marketplaces, the part that was supposed to replace the firm, are shrinking right now. Upwork’s volume in the second quarter of 2026 was 4% below a year earlier, and its CEO put that down to a “reduction of lower-value and highly automatable work.” Fiverr lost 22% of its buyers over the same year. Two companies don’t make a trend, but in Upwork’s own reading, the work that left is the work software can do.
Listed companies in the US
Down 52% from the 1996 peak.
New business applications a year
2025 was the record.
US businesses in 2023
Five in six have nobody on payroll.
Average headcount of a new unicorn
Companies valued at $1 billion for the first time that year.
Four counts, each on its own scale. Sources: the World Bank and the World Federation of Exchanges; the Census Bureau; PitchBook, via Bloomberg Opinion, April 2026.
What’s cheap in 2026, and what isn’t
Back to my afternoon. The work part of it is Coase’s employee, someone who obeys within limits, and in 2026 you rent it by the token. Its limits are its instructions and the tools you hand it. Its rate for a given level of skill keeps falling: output at GPT-3’s level cost $60 per million tokens in 2021 and $0.06 in 2024. Epoch AI estimated in September that the price of a given level of capability has been falling about 47% a quarter since 2023.
It also keeps a log of everything it did, which is the part I find strangest. Alchian and Demsetz’s whole problem was that you can’t see who did what in a team. With an agent, you read the transcript. It won’t tell you whether the work was any good, but it tells you exactly who did what.
My own setup reads like a job description. A settings file says which commands Claude may run on its own and which need my say-so, and that’s where this afternoon’s fork bounced. Those are Coase’s limits, written up as config.
So the agent is Coase’s employee in every economic sense and in no legal one. It takes directions within limits, works for a falling rate and keeps a record of every task, with no contract, no salary, no benefits and no hiring. That’s the employee without the employment.
- Inside
- Managers, memos and the telephone. Each extra transaction costs more to direct than the last.
- Market
- Finding a price, negotiating and writing a contract for every job.
Coase’s two curves, drawn from his description rather than from data. Inside: the cost of directing one more transaction, which rises as the firm grows. Market: the cost of buying it. Where they cross, the firm stops. The 2026 shapes are my guess, not a measurement.
Now run the five shortages through an agent. Knowing who did what goes to its log, and coordination at a distance has mostly been software’s job for years. Three don’t move. Kreps’s name still needs someone to answer for it, which in practice means a signature. Knight’s uncertainty still needs someone who can lose. Williamson’s hold-up is now about context. Add a fourth that none of the five covers, credit, and you have the part of my afternoon that waited for me.
The signature. The law still only sees people. A partnership in the US is “an association of two or more persons,” in words that go back to 1914. Courts have ruled that an AI can’t be an inventor or an author, and in April Tennessee passed an act saying a “person” “does not include artificial intelligence.” So someone with a name has to stand behind what an agent does.
In February 2024 a tribunal in British Columbia ordered Air Canada to pay CA$812.02 to a passenger its chatbot had misled about bereavement fares. The airline had argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal said Air Canada “is responsible for all the information on its website.” On blogs.hn, the one who answers is the name in the pull request’s title.
The EU’s new product liability rules go further. From 9 December 2026 they treat software, AI included, as a product, with no-fault liability for whoever made it. Either way, the one who pays is a person or a company, never the agent.
Credit. Banks lend to persons, and an agent isn’t one.
Uncertainty. Knight’s point survives untouched. An agent can estimate the odds. It can’t lose its house. If this site never finds its readers, I’m the one who paid for the servers.
Context. What a firm knows that outsiders don’t is the new hold-up asset, and it’s the scarcity my last essay ended on. An agent can read whatever a firm has written down. It can’t read what the firm knows and never wrote down, and that part only travels with the people.
That might explain a number that cuts against the idea that agents favour small firms. This spring the Census Bureau found 37% of firms with 250 or more employees using AI, against under 20% of firms with four or fewer. My guess is that big firms have more written down to point an agent at, though it may just mean small firms are behind.
Then there’s the fixed cost of a human employee, which hasn’t moved. Benefits were 31.6% of what US employers paid for an hour of civilian work in June 2026. An agent comes with none of that.
Put the cheap part next to the expensive one and you’d expect companies to form faster than ever and hire slower than ever. The broad numbers lean that way. Carta reports that 36% of the startups founded on its platform in 2025 had a single founder, up from 18% in 2016. Lovable passed $400 million of annual recurring revenue in February with 146 full-time staff.
Five bets, and what would prove me wrong
Same rule as last time. A theory that only explains the past is cheap, so here are five predictions, how sure I am of each, and what would change my mind.
1. Firms shrink in headcount, not in scope. Coase’s telephone effect wins. Agents cut the cost of organising inside faster than the cost of using the market, because inside is where the context and the log are, so a firm covers more ground with fewer people. (An agent you direct counts as inside. A finished job you buy from a freelancer or an agent marketplace counts as market.) Gillian Hadfield and Andrew Koh put the theory in one line last year: the obstacles that stop firms growing without bound “seem intrinsic to humans but not to AI.” The unicorn headcounts (with this year’s group still incomplete) and the shrinking freelance marketplaces are the early evidence.
How sure: fairly, on early numbers. What would prove me wrong: the share of work bought as finished jobs, from freelancers or agent marketplaces, rising.
2. The first “one-person” billion-dollar company will be a firm by another name. In late 2023 Sam Altman said his group chat of tech CEOs had “this betting pool for the first year that there is a one-person billion dollar company.” I think the winner will sit on top of freelancers, a law firm, an accountant and a bank, because the signature, the credit and the uncertainty still need persons. The closest case so far is Medvi, a telehealth company the New York Times profiled in April. It made $401 million in sales in its first full year with two staff, the founder and his brother, and its doctors and engineers are contractors.
How sure: half true already. What would prove me wrong: a verified billion-dollar valuation for a company with one human and no contractors, by 2030.
3. Hiring becomes the last resort. The next task goes to an agent before it goes to a new hire, because the agent’s wage fell and the employee’s fixed cost didn’t. The best evidence is Stanford’s “canaries” research. By June 2026, employment of 22-to-25-year-olds in the most AI-exposed jobs was about 19% below where it would otherwise be, up from the 13% the same team reported in August 2025. It came “through reduced hiring of young workers rather than increased separations.” On the other side, the Yale Budget Lab still finds no significant economy-wide effect.
How sure: early, and contested. What would prove me wrong: small firms that adopt agents adding staff faster than the ones that don’t.
4. Context becomes the boundary of the firm. A firm is what its agents can read and outsiders can’t. The evidence is faint, and a little morbid: context already has a price once a firm is dead. Brokers sell dead startups’ Slack and Jira archives for $10,000 to $100,000, and in August Google bid $10 million for Spirit Airlines’ emails and Teams messages at a bankruptcy auction. Nobody has built a market for the context of a firm that’s still alive.
How sure: a hunch. What would prove me wrong: markets for context that work across firm lines.
5. Liability becomes the product. The scarce thing is how many signatures a day a human can stand behind, and insurance for agent output becomes a product. It already has: Armilla’s Lloyd’s-backed cover for AI failures rose to $25 million per organisation in January. Courts are pricing the signature too. A database of court cases with invented AI citations had passed 2,225 by late October, and since June every lawyer who signs a filing in Florida vouches that the authorities cited exist.
How sure: some evidence. What would prove me wrong: liability shifting fully to the model vendors, so that buyers stop needing a person to stand behind the output. The EU’s rules from 9 December are the first real test. They put the liability on whoever made the software, which is still a company and not an agent, but it isn’t the person who signed for the output, so they lean against me.
1Firms shrink in headcount, not in scope
A newly minted unicorn averaged 1,128 employees in 2023 and 323 so far in 2026. Upwork’s volume fell 4% in a year; Fiverr lost 22% of its buyers.
Would prove me wrong: The share of work bought as finished jobs, from freelancers or agent marketplaces, rises.
2The one-person billion-dollar company is a firm by another name
Medvi: $401 million in sales in its first full year, two staff, and outsourced clinicians and engineers.
Would prove me wrong: Someone reaches a $1 billion valuation alone, with no contractors behind the curtain, by 2030.
3Hiring becomes the last resort
Workers aged 22 to 25 in the most AI-exposed jobs are about 19% below trend, through less hiring. Yale sees no economy-wide disruption yet.
Would prove me wrong: Small firms that adopt agents add staff faster than the ones that don’t.
4Context becomes the boundary of the firm
Dead startups’ Slack archives sell for $10,000 to $100,000. No market exists for a living firm’s context.
Would prove me wrong: Markets for context that work across firm lines.
5Liability becomes the product
Lloyd’s-backed cover for AI failures reached $25 million per organisation in January 2026; courts have logged 2,225 cases of invented citations.
Would prove me wrong: Strict liability shifts fully to model vendors. The EU’s product rules from 9 December 2026 are the first test.
Dots: how much evidence there is today, out of four.
The five bets. The sources are in the text above.
Write down why you’re a firm
My last essay ended by asking you to add two lines to your architecture decision records. This one asks for two more, on the first page of your company’s decision record, your partnership agreement, or the README of a company of one.
Why inside: ___ (context / liability / trust / uncertainty)
Revisit when: ___The first line is what your company supplies that an agent and a contract can’t. The second is what would have to change for that to stop being true. If you can’t fill in the first line, the market can probably do your job cheaper.
Here are mine, for the one-person operation behind this site.
Why inside: my name on everything that goes out, and the call on what deserves to.
Revisit when: a directory, a reader or a court takes an agent's word with no person's name under it.Post yours in the comments. I’d like to read them.
A thought to add?
A good question, another perspective, or something this made you think about. There’s room for it here.
Loading the conversation…