Peter Diamandis introduced a paper to the panel and set up the question behind it. In 1937, the economist Ronald Coase asked why companies exist at all — a question, Diamandis said, that won him the Nobel Prize. Coase's answer was about cost: transactions are expensive. Finding the right people, negotiating with them, writing and enforcing contracts all take time and money. Past some point it is cheaper to hire employees and keep the work inside the company than to negotiate every task on the open market. In that account, a company is a workaround for transaction costs.
The new twist is what happens if those costs collapse. MIT and Harvard researchers, Diamandis said, had asked what follows when AI agents — software that can search, compare prices, negotiate and transact on your behalf, at very large scale — make transactions nearly free. They call it the Coasean singularity.
Salim Ismail, who has argued versions of this on the show before, gave the paper a backhanded endorsement: "Absolutely dead on, a little late, but totally dead on."
Where the framework comes from
One panelist checked the date and noted that this was research from last year rather than something newly published. The presentation behind the phrase is a September 2025 NBER workshop talk by John Horton with Peyman Shahidi, Gili Rusak, Ben Manning and Andrey Fradkin, which looks at agents as participants in markets and starts from the buyer's side rather than the firm's.
Its running example is ordinary: buying a grill. Even that transaction involves specifying what you want, gathering information, comparing offers, haggling and then dealing with delivery or a dispute. Agents can cut the human time each of those steps takes. The framework expects better matching as a result, and longer negotiations, because software does not get bored. It also expects problems that do not disappear with the effort of shopping: attempts to manipulate agents, prices that become harder to see, and congestion when applying to everything is cheap and everyone's agent does it. And it sketches market designs that were impractical when humans had to sit through them — rapid multi-round matching, structured windows for negotiation.
These are proposed mechanisms and predictions, not measured outcomes. What happens depends on how agents are rewarded, how platforms are designed, and whether the parties bear the costs they impose on everyone else.
Ismail's version: the firm as a protocol
Ismail traced his own reading back to Exponential Organizations 2.0 in 2023, where he and his co-authors argued that Coase's law was already breaking, without understanding the full implications. What he noticed then was that companies had begun doing their most important work outside their own boundaries. Uber's mission-critical function is matching a driver with a passenger, and that does not happen inside Uber's organizational boundary — it happens out in the wild, and enabling it with technology is what let the company scale. XPRIZE goes to teams all over the world to do its innovation. TED uses its community to scale.
Agents, in his account, change the game rather than extend it. When coordination costs fall toward zero and a thousand or a hundred thousand agents can do "crazy amounts of capability" outside the organization, AI "doesn't just automate the firm. It attacks the economic reason for which firms exist and the shape of the firm." What he and his collaborators have been working out is what that shape becomes: a firm as a protocol, and a community of agents and human beings attacking economic opportunities, marketplaces or specific problems. The organization, as he put it, dissolves from a human, hierarchical, centered model into something else — the biggest shift, in his view, since the Industrial Revolution — which is why the governance of those agents is now part of the work.
Pushed on the obvious implication — if high transaction costs argue for large firms and low ones for small firms, does firm size head toward one person or fewer? — Ismail said he saw no limit. He pointed to what he called the Argentinian model, where agents could run the firm outright, and said there is no reason a legal entity cannot be owned by other agents.
What is left: the fiduciary wedge
That is where his answer stops short of dissolving the company entirely. In a paper with Ted Shelton, Ismail coined the term fiduciary wedge for the residue that still needs a legal entity: liability, fiduciary duty, ownership of proprietary data, the learning loops that accumulate inside an organization, brand, and the entity as a container for purpose — what he calls a massive transformative purpose. His analogy is the special-purpose vehicle in investing, where several people join a legal container to invest in one thing. The container survives; the reason Coase identified for it — coordination and execution — is what he expects to disappear.
The countervailing pull toward bigger firms
Alex pushed back with a force running the other way. If frontier capabilities get walled off — if OpenAI, say, uses internal unreleased models to attack grand challenges in mathematics while the rest of the economy has no access to them — then the incentive is to be inside the lab. He credited the point to a friend at OpenAI and others: that would argue for firms growing larger and larger, so that more people sit within reach of those internal capabilities. Exactly the opposite of the Coasean story.
Ismail allowed the case but called it an edge case. Open models, in his view, put general intelligence and agents in most people's hands across the board. A frontier lab could already be running its own hedge fund strategy on a model that beats the market, but he treated that as niche and temporary. His larger worry was elsewhere: if recursive self-improvement — AI systems improving the systems that build AI — actually arrives, "the concept of an economy starts to erode and dissolve itself," and the question has to be posed in a different model entirely.
Diamandis reached for the old riddle about an irresistible force meeting an immovable object: a Coasean frontier lab with superintelligent capabilities on one side, an agentic economy pushing transactions out to the edge on the other, and nonhuman corporations in the mix. Not obvious, he said, where that adds up. One reply on the panel took the argument to its endpoint — the scenario Dario Amodei has described, where a single company like Anthropic is effectively all of private enterprise — and judged it unlikely, while agreeing that the ordinary meaning of "an economy" would be gone by then, leaving open where value is created and stored.
"The economy needs a bit of friction"
Emad Mostaque took the other side of the transaction-cost claim. "The economy is like 1% inspiration, 99% perspiration," he said: you do not need a polymath doing your taxes or selling widgets. Even where a model produces the recipe, the work is in following through. Most transaction costs, on his reading, are friction — and "the economy needs a bit of friction." It is the relationships you build and the other things that are not instant. Nothing spreads without barriers; it is not the case that the best product always wins, "or we'd all be on beta-max."
His forecast is therefore smaller than Ismail's but not a solo act. He does not expect the one-person company. He expects the ten-person company, and more collectives of companies working with digital and physical humans to solve problems that deliver value, because organizations get unpleasant past twelve people and again past a hundred and fifty, where disorder creeps in and teams go misaligned. What does have to change, in his view, is the underlying economics: from scarcity-based to abundance-based, as bits and eventually atoms become rearrangeable.
Both directions at once
Dave closed by stepping back from the academic frame. Mercor, he said, now has something like fifty to a hundred thousand individual actors who are effectively little companies, in India and now Brazil too. Someone could build a marketplace around exactly what Ismail describes, handling the lingering artifacts such as employment law in different countries, and make a fortune riding the trend. Concurrently, Elon Musk is building the most vertically integrated company the world has seen, with supply-chain control running down to raw sand turning into chips.
Both are happening in the real world at the same time, Dave said, and for good reasons in each case — which leaves the question the workshop framework poses still open. Cheaper transactions do not by themselves say how large a firm should be; they change one cost among several, while liability, data, relationships and access to the best models pull in their own directions.