AI has arrived at our problem

Dear friends and collaborators,

Last week, on 29 October, Alex Imas, Andrew Koh and Shane Legg of Google DeepMind published Organizing Intelligence, the most careful model yet of when AI transforms the economy. Their answer deserves to become common sense: intelligence is no longer what is scarce. Coordination is. The single AI agent, they write, is ”the wrong economic object.” A model can beat every person one-on-one and still lose to a well-organized team. In their words, ”competence reshapes the economy gradually, whereas coordination abruptly transforms it.” Once AI agents coordinate well enough, returns to scale turn convex and the Coasean ceiling, the limit on how large an organization can grow, breaks.

We agree with the diagnosis. What the economy is, is about to be reconfigured.

But we question the one thing their analysis takes for granted: the firm. Why should the firm still be the place where all that coordination lands?

In their model the firm is the only organization that competes, so supercooperation tips markets ”toward large, concentrated AI-heavy firms”. That follows from assuming the firm is the unit of economic coordination. Ronald Coase’s thesis, in 1937, was that firms exist where coordinating inside them is cheaper than coordinating through the market. But AI lowers the cost of coordination from every position, not only from the top of a firm. As it does, more of what used to take a firm comes within reach of anyone with an agent. AI is sold to us as the replacement of the job. It is more the replacement of the firm.

So our wager is different: economic coordination beyond the Coasean ceiling can take an internet-native network form. A network of principals and their agents, coordinating agent to agent through shared protocol rather than under one roof. The internet did this once for publishing and for software: its increasing returns went to open protocols, not to walled services. What the internet never got was an economic protocol native to it. That is why the platform startups were able to capture it. And they are attempting it again, now at AI scale. What we need is the economic network harness: a protocol-equipped environment that composes an agent-to-agent economic medium. The rest of this letter is why we think so, and what we have built toward it.

Pivot to AI? Not really. AI arrived at our problem.

Our economic agent architecture and our thesis of distributed economic agency predate the current AI-agent cycle. Since 2016 we have been specifying 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. In that work, ”agent” always meant the participants of a networked economy and the processes acting for them. This year AI arrived at that problem, in two ways.

First, AI exposed the economy as a medium. The public conversation about AI is mostly about jobs. But the job is one part of a larger arrangement: the firm, money, the price, the platform, the account. These have been our economic coordination interfaces. They made complex coordination workable when information was expensive. Agents now reveal them for what they are: interfaces built for the informational constraints of another era. When the conditions change, interfaces get redesigned. Read closely, most critiques of AI are critiques of how AI is deployed economically: trained by extracting the commons, sold to firms as a replacement for the people who work in them, with authority staying where it was. AI is not the problem. The economic harness around it is.

Second, AI changed what is feasible. Large language models did not originate our thesis. But they changed its feasibility twice over: as an intelligent interface for protocol-mediated agents, and as a multiplier on what a small expert team can build. They lower the cost both of building the architecture and of living in it. This is the time of the protocol, and of the network.

Three barriers, opened

Three barriers have stood between a specified postcapital economy and a running one. This year all three moved.

• Technical. Executing an economy as a protocol, not only modeling it, is a distributed-systems and formal-methods problem across many disciplines at once. It used to take years and a roomful of very expensive specialists. AI raises a small team’s cross-domain output enough to make the protocol’s formal design tractable: understandable and executable. Formal methods are now how we specify. Our own expertise and verification stay at the center.
• Organizational. A decentralized, non-hierarchical network is informationally expensive. Our organizational experiments taught us that hierarchy exists for a reason: it is a complexity-management function, necessary under certain informational conditions. Now AI can absorb much of that complexity behind an interface in plain language. You state your goals, your agent proposes the network actions, and the rules stay inspectable. That is the interface our new economic interaction language was missing. What better interface to a new economic-organizational language than one that speaks it and acts in it fluently?
• Funding. Building an economy beyond profit takes capital, and capital comes only for profit. It looks like a paradox. AI turns it into an opportunity. When agents become everyone’s interface, the agent-to-agent economy needs an economic medium of its own, and whoever finances that medium early holds a position in all of it, not in one lab. The bridge we have designed is how capital takes that position: a mechanism that makes financing the network profitable for capital, while staying structurally unable to impose profit logic on what it funds. Capital comes in and funds the production of distributed economic network computation as a concrete, usable utility. Agents multiply how much of that utility every principal can produce and use, so the more agents work through the medium, the more the network produces. The bridge mechanism turns that growth into capital’s upside, and the funder’s grammar never becomes the funded network’s law. The network’s course stays its own.

What a small team can do now

In autumn 2025, after the Collaborator Round, we made a bet: before raising the Seed Round, understand fully how AI changes what we are building, then build with it.

The bet paid off fast. By extending AI reasoning with executable virtual environments and formally verified code, we found ourselves with a new development machinery. Specification, formal analysis, simulation and implementation now run as one process, with agents working alongside us across all of it. That machinery is also what opened the funding barrier: the bridge, once a set of functions on paper, is now a formal model whose safety properties are accounting identities, holding by construction, and it is computationally formalized and simulation-tested. Testnet deployment and an independent audit are next.

We also started coordinating through the medium we are building, and designing our organization as a filesystem. Since April: a friend-to-friend shared vault that contributors join by invitation, with no server and no accounts. Since May: two agent sub-organizations helping run our lab and our knowledge vault, each working for named people and steadily extending what we can jointly handle.

A small team with this machinery can attempt what used to take an institution.

But we need to keep learning: we also put together an AI harm mitigation guide for small teams.

The economic network harness

Here is what came out of the year. Agents already have harnesses for coding, for research, for memory. None of them is economic. An agent can write your code, but it cannot hold a commitment, extend credit, share in a surplus or form an economic relation on your behalf, within limits you set.

That is the economic network harness: a protocol-equipped environment in which any agent, yours and not only ours, manages its principal’s economic relations, with economic state, authorized operations and scoped authority. AI gives agents intelligence. We give everyone an economic network agent. With one, anyone can act as a principal: set the purposes, let the agent choose the means, and start an economic space as easily as starting a repository.

This is also how it spreads. The platform won the last round by owning our interface. When your interface is an agent, the interface we can distribute is the harness itself. Booting the harness is joining the network: start it up, and you are in, with an agent that knows how to act there.

The harness is in its first phase: its pieces run among contributors today, and its first distributable version, for a first ring of invited participants, is close. If you want to be among its first test users, send us a message: hello@ecsa.io.

What comes next

We are opening the Seed Round with a new positioning and a new deck, The economic network harness: an agent-to-agent economic medium. For the next three months these bulletins walk through its argument one step at a time, so you can follow it, question it and tell us what you think. The first three steps:

• Part 1 · The stakes: Who gets to write the economic rules? What if that write permission were distributed?
• Part 2 · The harness: Why is AI sold as the employee’s replacement, when the arrangement it really replaces is the firm?
• Part 3 · The prize: What becomes possible when principals author their own economies?

Want to see the deck now, or the data room? Write us at hello@ecsa.io. And if you just want to say hi, we’ll love that too, on Telegram or Bluesky.

See you there!