Why the Firm?
What is the right economic object for judging AI's impact?
A reply to Alex Imas, Andrew Koh and Shane Legg
by Jorge López and Akseli Virtanen
On 29 October, Alex Imas, Andrew Koh and Shane Legg of Google DeepMind published Organizing Intelligence, with a companion essay, The Economics of Intelligence. The paper develops a model, the most careful yet, to analyse when AI transforms the economy. Their argument deserves to become common sense. Intelligence is no longer what is scarce; coordination is. A model can beat every person one-on-one and still lose to a well-organized team. "Competence reshapes the economy gradually, whereas coordination abruptly transforms it." And so "the individual AI agent is the wrong economic object."
We agree with the diagnosis. But we would like to take the argument one step further. They put the organization in the agent's place: "the right question is whether an organization of AI agents can surpass an organization of humans." Past a coordination threshold, AI can "break the Coasean ceiling and trigger a sudden phase transition toward a giant AI firm."[1] We think the organization is not the right economic object either. The network it belongs to is.
Two kinds of coordination: one principal and its agents vs. a network of principals
An AI agent is a model in a harness. However capable it is, it acts for someone. Agents carry authority and pass it on; principals hold it, and only principals determine what their agents may commit them to. That is the older meaning of the word, from agency law, and it is the part that today's usage of "agent" forgets. An agent always has a principal.
Seen this way, a firm is one principal with its agents. Coase's workman moves between departments "because he is ordered to do so": in the language of agency theory, he acts for the firm. In the paper's model, workers and AI copies both act as agents of the firm, and the objective the agents execute is the firm's. We use the term economic agent for this: the interface through which a principal acts in the economy, whether it is a role a person plays, an AI, or both. People have economic agents. They are not them.
But no principal acts alone. The economy is a network of principals, and a formal one: every receivable in one firm’s books records a claim on someone else, and every contract binds one principal to another. A firm is one node in that network. When the paper has firms compete, it studies the nodes and leaves out the links between them.
Those links are coordination too, though of another kind. What sets a firm apart from the economy around it is how authority runs. In a firm, it starts at one principal and runs one way, down to its agents. That is direction. Between principals it runs both ways, each binding the other and bound in turn. That is contract. AI agents make both kinds of coordination cheaper, but a firm runs on the first alone. To judge what AI does to the economy, we have to look at the network, where both kinds of coordination meet.
Coase cuts both ways
Coase asked why firms exist if markets work. His answer was that using the market has a cost: "discovering what the relevant prices are," "negotiating and concluding a separate contract for each exchange transaction," and then seeing it through. A firm replaces that series of contracts with internal direction, to save on that cost. Its boundary sits where organizing one more transaction inside costs as much as contracting for it outside: between the cost of direction, within one principal, and the cost of contracting, between principals. Whether an invention makes firms larger or smaller, he wrote, "will depend on the relative effect on these two sets of costs."
The paper models how agents extend what one principal can direct. They execute the firm's objective, and as they grow more capable, direction gets cheaper and the firm grows "from within." For the second cost, the authors point to work on AI agents that remove "market search and contracting frictions." We think the second cost also belongs in any model of AI's impact on the economy. Agents extend what principals can agree on, just as they extend what one principal can direct; in both cases they lower the cost of coordination, and only the way authority runs differs. A model that lowers one of Coase's two costs will see only one of his two outcomes.
Between principals, authority runs both ways, so every step has to be agreed rather than ordered, and agreeing takes far more information than ordering. The firm is how the economy has saved on that cost: it is an economic coordination interface that lets a network of participants grow by making most of its links run one way. But agents now lower that cost across the whole life of a contract, from defining and negotiating the terms to putting them in force and fulfilling them. At the network edge, one person with agents can do what took a modern firm. That person is still one principal directing their own agents, but such principals will multiply, and more of the economy's coordination will run between them. On 9 November, the Wall Street Journal reported how "20-Somethings Are Ditching Entry-Level Jobs to Become Their Own Bosses".[2] AI is sold as the replacement of the job, but it is at least as much that of the firm.
How much coordination between principals has been held back by its cost alone? Finance gives some measure. It is the richest coordination principals have with each other: credit, insurance, contingent claims, terms fitted to a relation. It costs enough that most principals never get near it and settle for standard terms, reduced to a price and left to default law. The economy still pays heavily for the finance there is, and demand that holds at such a price suggests how much more there would be at a lower one. Agents bring that kind of contracting within reach of any principal.
Which way will the economy go?
If AI lowers both costs, which way will the economy go? The paper's own warning applies to both: "an economy near its tipping point can be reorganized by a coordination improvement too small to register on any benchmark." The warning says nothing about direction. A gain inside the firm tips the economy toward a few giant firms. A gain between principals tips it toward a denser network of independent ones. Which comes first depends on what gets built. Today's models and agent tooling are built for the firm, because firms are who pay for them. So the firm is the default and that says more about who pays for the tools than about which side has more to gain.
But the two outcomes are different economies. Coase, quoting D. H. Robertson, described firms as "islands of conscious power in this ocean of unconscious co-operation."[3] A few giant firms would be all island and no ocean: planning at the scale of the whole, which is not what we usually mean by an economy. What we value in an economy, many principals finding together what none of them could plan, belongs to the ocean.
They also differ in what they can compose. The paper has a name for what difference is worth. Copies of one model "overlap in their jaggedness," sharing strengths and blind spots alike, while people differ, and sorting different people into teams creates mosaic intelligence, which "can out-compete more competent AI." An economy is mosaic intelligence at its richest, with more knowledge and more purposes than any firm, however large, can hold under one direction. A firm composes its employees by subordinating them to that direction, and each belongs to one firm. Principals who coordinate as principals can take part in many compositions at once, each serving different ends. If new value comes from new combinations that is where the larger gain lies.
This is the side we have chosen to build for. The firm solves a real problem: composing many people's agency into one capacity to act. It solves it by subordination, which is how much of labour law defines employment. We work on composition without subordination, so that everyone, already a principal, can act with others at a scale only firms have reached. Inside a firm, direction holds the composition together. Between principals, agreement has to, and agreement needs shared meaning, verifiable records passed agent to agent, explicit commitments, and authority that can be traced to whoever granted it. It needs a different harness. Since 2016 we have worked on specifying the protocols that provide that: the protocols through which economic agents establish identity and authority, account for activity, form commitments, constitute credit and stake, organize production, govern their relations and share in the surplus they create. This economic network harness, which composes an agent-to-agent economic medium, is the first implementation of that work, and it is in its first phase.
An invitation
Whether principals can coordinate at the scale direction allows is an empirical question. The paper already suggests how to measure it: benchmarks, it says, should track "how output scales with the number of identical AI copies at fixed compute per copy," across model generations. We would add one dimension. Hold compute fixed, and measure the same scaling for agents acting for different principals, with different objectives, coordinating through shared records and explicit commitments. How far does it fall short of coordination under one direction, and what closes the gap? Does the gap narrow as models get more capable? How much does the difference between principals add back?
Other questions come before any benchmark. How much coordination between principals is waiting on its cost? When every principal can write contingent terms, where does the risk go, and who ends up holding it? What has to be true of the records between principals for that to be safe? We are already working on the answers.
We share the authors' starting point: the economics of intelligence is the economics of coordination. Whether AI-scale coordination and its benefits end up inside a few firms or in a medium everyone can join is being decided now, mostly by what gets built.
We would welcome working on it together.
- Ronald Coase, "The Nature of the Firm," Economica 4, no. 16 (1937): 386–405. Coase asked why firms exist at all if markets coordinate production. His answer was that using the market has costs, which a firm avoids by directing work instead. A firm grows until organizing one more transaction inside costs as much as contracting for it outside. The paper's "Coasean ceiling" is that limit on firm size, which in its model comes from how coordination degrades as a firm adds members. Past a threshold of AI coordination, in the companion essay's words, "bigger becomes better at every scale, and firm size is only constrained by the size of the market." Coase received the Nobel Memorial Prize in Economic Sciences in 1991, in part for this paper. ↩︎
- "20-Somethings Are Ditching Entry-Level Jobs to Become Their Own Bosses," The Wall Street Journal, 9 November 2026, https://www.wsj.com/lifestyle/careers/20-somethings-are-ditching-entry-level-jobs-to-become-their-own-bosses-168d1f48 ↩︎
- Coase (1937), p. 388, quoting D. H. Robertson, The Control of Industry (1923). Robertson's image continues: "like lumps of butter coagulating in a pail of buttermilk." Coase uses it to pose his question: if prices coordinate the economy, why do these islands exist at all? ↩︎