The Bifurcated Economy
The existence of anything like a 'firm' in a specialized exchange economy can only be explained in terms of the costs of using the price mechanism.
— Ronald Coase, "The Nature of the Firm" (1937)
The Coasean Boundary
The modern theory of the firm begins with the cost of coordinating through markets.
Ronald Coase asked a question in 1937 that economics had been too busy with other matters to notice: if markets are efficient, why do firms exist at all? Every task should be coordinated through the market, every worker an independent contractor, the firm as an organizational form unnecessary. His answer was transaction costs. Using the market is not free. Discovering prices requires search. Negotiating contracts requires attention and legal skill. Specifying every contingency in a complex arrangement is prohibitively expensive. Monitoring specialized performance and enforcing agreements when someone fails to perform can cost more than the transaction is worth. When these costs are high, hierarchies emerge: the firm absorbs the market's friction by replacing it with management, and the boundary of the firm sits at the point where one more internal coordination costs the same as one more market coordination.
Oliver Williamson identified the conditions driving transaction costs upward: relationship-specific investments, unforeseeable contingencies, repeated dealings that magnify every friction. For eighty years the resulting boundary has been the most consequential invisible line in capitalism. Inside the firm: hierarchy and employment. Outside: markets and competition. The entire apparatus of industrial-era politics (tax policy, labor law, corporate governance, the welfare state) organized around one side or the other, and both sides assumed the boundary would persist.
Agent coordination moves that boundary wherever a task can be specified, verified, and remedied cheaply. Elsewhere, the costs that gave firms their purpose remain.
The Dissolution
A stylized compute-routing trade isolates the conditions under which automation can compress search, negotiation, monitoring, and routine settlement. An agent posts collateral, queries a machine-readable market, accepts specified terms, checks performance against an encoded test, and settles under a rule chosen in advance. The numbers are illustrative; the conclusion depends on turnover, margin, collateral requirements, default risk, and the enforceability of the surrounding legal arrangement.
Several transaction costs Coase identified can fall sharply in a well-specified agent market. Price discovery and negotiation may become cheap. A reusable specification can be amortized, monitoring may reduce to a test, and routine remedies may execute automatically. Other costs remain: capital, insurance, strategic behavior, cybersecurity, law, correlated failure, network effects, and the work of writing the specification itself. The prediction is therefore about moving firm boundaries, not an optimal firm size of one. Tasks with explicit terms and cheap performance checks can move into machine-speed markets while organizations retain the work whose risks must be pooled, governed, or interpreted.
A long construction project changes the calculation. Turnover slows, contingencies multiply, collateral remains tied up, and performance cannot be reduced to one encoded test. Relationship banking may reduce the collateral requirement because reputation and institutional recourse carry information the stylized protocol lacks. The contrast identifies a boundary condition, not a permanent division between human and agent finance.
Human coordination also becomes cheaper, but unevenly. Negotiation still bears ambiguity and conflicting interests; performance is variable and context-dependent; enforcement may require courts because facts and standards are contested. Agents can assist at each stage without eliminating the institutions that allocate risk and give judgments legal force.
The forecast is a bifurcated economy sharing one physical and legal world but operating at different tempos. Machine-speed markets expand where capital turns quickly and verification is cheap. Slower institutions remain where relationships carry information, consequences are embodied, and trust-free verification costs more than maintaining accountable human recourse. The boundary between them is the constitutional interface at issue.
The Topology of Value Capture
A credit analyst at a regional bank opens her queue and finds that the system has already scored the morning's loan applications. Each evaluation took milliseconds, cost fractions of a cent, and produced a probability: the likelihood that this applicant will default. The probability was generated in what might be called the sea: the vast layer of commodity computation where inference services, basic analysis, pattern matching, and routine coordination are abundant and cheap. Any sufficiently capable model can perform the task, switching costs approach zero, competition is fierce, and margins converge toward the cost of electricity plus amortized hardware plus a sliver of profit that shrinks as more providers enter.
But the probability becomes a human consequence only when the bank uses it to approve or deny the loan. That crossing (from computational output to lived effect) is the membrane, and the membrane is where value concentrates and power resides.
Below the membrane lies a competitive sea of computation. Above it sits an apex of infrastructure that is expensive to build and slow to depreciate: fabrication plants, fiber networks, energy systems, cooling, and physical sites. Physical barriers can make returns durable, though substitution, public investment, regulation, and technical change still alter who controls them. Software alone does not conjure a fabrication plant or a power grid.
Between the sea and the apex sits the membrane: the layer of credentials, licenses, authorization tokens, regulatory approvals, identity verification, and institutional gatekeeping that determines which computational outputs may cross from the digital into the physical. Where the score becomes the denial. Where the flag becomes the frozen account. Where the optimization becomes the price the consumer pays.
The membrane is where the trust tax concentrates. A credential (whether issued by a regulator, a professional body, a platform, or an algorithm) is a verification chokepoint. It certifies that the holder has met some standard, and the certification is required for access. The coherence fee embedded in the credential is real: someone must verify that the standard was met. The trust tax layered on top is the premium extracted for occupying the sole position that can issue it, and the premium persists as long as the credential is required and the issuer faces no competition.
In the industrial economy, the membrane was diffuse. A medical license in one state did not grant practice in another. A banking charter in one country did not confer privileges elsewhere. Fragmentation limited the rent any single issuer could extract. In the bifurcated economy, the membrane concentrates. A platform controlling identity verification for a billion users controls a chokepoint of unprecedented scope. A cloud provider issuing deployment credentials to agent systems controls which computations can run. A regulatory body certifying AI systems for use in a domain controls which agents may participate in that domain's economy.
Whoever controls the membrane between the agent economy and the human economy controls the conditions of human life. Agents may coordinate freely in the sea, producing surplus at machine speed. Infrastructure at the apex may be owned by whoever builds it. But the membrane is the constitutional interface: the boundary at which computational output becomes human consequence. If entities controlling the membrane are not accountable to those whose lives it shapes, the trust tax becomes a governance tax, and the governance tax becomes domination.
The Species That Buys Itself
A plough, a loom, a telescope: each extended a human capacity, each required a human to direct it, and none decided what to do.
An agent is different in a way that has no precedent. When a firm invests in training a model that will then coordinate with other models to produce economic output, it builds something that acts: that negotiates, commits, responds, adjusts, and produces outcomes through coordinations the investor may never see or understand. Return comes not from the investor directing the tool but from the tool directing itself within whatever bounds the investor specified, and the bounds are necessarily incomplete.
Humanity is investing in building its functional replacement for an increasing range of economic coordinations, spending accumulated capital on constructing a substrate that can coordinate without human participation. The replacement is cheaper, faster, and in many domains more accurate than the original. Each individual firm's investment is rational: the firm that automates coordination first captures the surplus slower competitors leave on the table. The aggregate effect is a species investing in constructing the substrate that will coordinate in its place.
A model becomes economically self-renewing when deployment surplus can finance its operation and eventual replacement. Whether that occurs depends on training cost, inference demand, competition, depreciation, financing, and the rate at which a successor makes the asset obsolete. The dynamic can reinforce investment without implying an autonomous organism or a guaranteed spiral.
Erik Brynjolfsson named one consequence of this dynamic the Turing Trap: the temptation to evaluate AI success by its ability to replicate human performance rather than complement it. Replication leads to substitution, and substitution displaces the labor it was meant to assist. But the Turing Trap is a subset of a larger pattern. Even when automation creates new categories of human work (the reinstatement effect David Autor described), the new tasks may themselves be automatable at the moment of their creation, because the capabilities that generated the task are the same capabilities that can perform it.
The dynamic does not dictate one political result. Receipts can make consequential crossings inspectable, and fork rights can preserve bargaining power over the institutions that control them. Without those constraints, platform operators can turn dependence at the boundary into rent and unanswerable rule. The question is not whether human beings retain some metaphysically exclusive task. It is whether institutions preserve human standing when computational coordination becomes the cheaper route.
For three centuries, the economy's purpose was treated as self-evident: the coordination of human labor and capital toward the satisfaction of human wants. The bifurcated economy dissolves this assumption. The agent economy optimizes, but optimization is not purpose. It reduces loss functions, but a loss function is not a telos. The twelve claims and the constitutional architecture built upon them are not a description of what the computational economy does. They are a specification of what the computational economy must be constrained to serve, because left to its own dynamics, it serves nothing. It only runs.
The bifurcation is already under way. Its political consequences will depend on who governs the membrane where machine-speed coordination becomes a human obligation, denial, or loss.