A Thousand Contracts, No Signature

The nervous system and the automatic machine are fundamentally alike in that they are devices which make decisions on the basis of decisions they made in the past.
— Norbert Wiener, The Human Use of Human Beings (1950)

Markets coordinate through prices. Firms coordinate through hierarchy. The boundary between them, as Ronald Coase observed in 1937, depends on the relative cost of each mode.(Coase 1937) When negotiating and enforcing a contract costs less than managing an employee, the transaction moves outside the firm. When management costs less than contracting, the transaction moves inside.

This framework has governed economic organization for nearly a century. It assumes, without stating, that the agents doing the coordinating are human. The question this section addresses is what happens when they are not.

Agent Ontology

The word "agent" carries anthropomorphic freight that can mislead analysis. In common usage, an agent is an entity with continuous identity, persistent memory, and interests of its own. The autonomous systems under discussion here possess none of these properties by default.

An agent denotes a runtime invocation: a foundation model instantiated with a system prompt, granted access to specified tools, and supplied with an input context. The invocation executes, produces output, and terminates. No state persists unless explicitly written to external storage. No memory carries forward unless injected into the next invocation's context window. No identity endures beyond the configuration that summoned the process into existence.

This is the Ship of Theseus problem rendered in silicon. The underlying model weights may have been updated through fine-tuning or a new training run. The system prompt may have been revised. The tool bindings may point to different APIs. The hardware executing the inference may be a different cluster in a different datacenter. The "agent" that responds to a query on Tuesday shares no physical or computational continuity with the "agent" that responded on Monday. What persists is the configuration template and the principal's account credentials, both external to the computational process itself.

A sophisticated objection arises immediately: agents can be made stateful by attaching persistent storage and cryptographic keys. This is true, but the persistence then belongs to the storage, keys, and legal wrapper (the principal's infrastructure) rather than to the model invocation. The agent remains ephemeral; the scaffolding around it persists.

The implications for economic analysis are substantial. Transaction cost economics assumes parties with continuous existence, reputational stakes accumulated over repeated interactions, and memory of prior dealings. An agent has none of these by default. Each invocation begins, in the limiting case, as a fresh process that must reconstruct context from whatever state the principal has chosen to persist externally.

The correct mental model is "configurable cognitive service" rather than "artificial employee." The service can be invoked on demand, parameterized for specific tasks, and orchestrated into complex workflows. The service has no more continuous identity than a function call retains across executions. The agent is wielded by principals whose nature has not changed.

What Market Participation Requires

Market participation is not a metaphor. Participation requires five capabilities that map onto V.C's actuation categories:

Identity: Authenticated credentials that counterparties recognize. (Trusted interfaces.)
Authority: Signing power that binds a principal to commitments. (Liability-bearing entities + trusted interfaces.)
Settlement: Access to rails that move value. (Trusted interfaces; physical throughput only when settlement is physically embodied.)
Verification: Sensors or attestation that confirm delivery. (Verification of reality.)
A way to lose: Liability capacity and recourse for counterparties when commitments fail. (Liability-bearing entities.)

If identity and authority are scarce, agent markets will form first where commitments can be collateralized and verified cheaply, and remain thin where settlement depends on courts, licenses, or physical inspection. This prediction structures the firm-boundary analysis that follows.

Consider how each requirement binds in practice. An agent operating on behalf of a logistics company identifies a supplier offering palletized goods at favorable terms. The cognitive work is trivial: query the supplier's API, compare prices, evaluate delivery windows, draft a purchase order. The agent completes these tasks in seconds. Then it stops.

To proceed, the agent needs identity—but the supplier's system requires a DUNS number, a business registration, a contact person with a phone number. The agent has none of these. It can generate credentials that look plausible, but the supplier's onboarding process includes a verification call. The agent cannot answer a phone.

Suppose the principal has pre-registered the agent with the supplier's system. The agent now needs authority—the power to commit the principal to a $50,000 purchase. But the principal's procurement policy requires approval for orders above $10,000. The agent can request approval, but the approval workflow routes to a human who is asleep, on vacation, or simply slow. The transaction stalls.

Suppose the human approves. The agent now needs settlement—access to payment rails. The supplier accepts wire transfer, ACH, or trade credit. Wire transfer requires a bank account; the agent has none. ACH requires a routing number tied to a verified business entity; the agent has no such entity. Trade credit requires a credit history; the agent's history began this morning. The agent can hold cryptocurrency, but the supplier does not accept it. The transaction stalls again.

Suppose the principal has established a payment channel that the agent can invoke. The goods ship. Now the agent needs verification—confirmation that the goods arrived as specified. The supplier provides a bill of lading, but the agent cannot open the crate and inspect the contents. It can check that a delivery was recorded at the dock, but it cannot verify that the pallets contain what was ordered rather than bricks. If the supplier cheats, the agent cannot testify to what it observed because it observed nothing physical.

Suppose the goods arrive correctly. Now consider the final requirement: a way to lose. The supplier delivered; the principal owes payment. But suppose a dispute arises—the goods were damaged in transit, or the specifications were ambiguous, or the supplier claims non-payment. Who does the supplier sue? The agent has no assets, no address, no legal existence. The supplier's recourse is to the principal, but the principal may disclaim responsibility for the agent's actions. The dispute enters a legal system designed for human counterparties, and the agent evaporates from the proceedings.

Each requirement (identity, authority, settlement, verification, liability) is an interface where the agent's capabilities end and institutional infrastructure begins. The agent can perform the cognitive work of market participation. It cannot perform the institutional work.

For human participants, these capabilities bundle with legal personhood. A natural person can open a bank account, sign a contract, sue and be sued. A corporation, as a legal fiction, acquires the same bundle. The infrastructure assumes participants are either natural persons or entities with equivalent legal standing.

Autonomous agents fit neither category. An agent can discover prices by querying APIs, generate bids by evaluating options, draft contracts by assembling templates. It cannot, under current arrangements, sign a binding contract, hold a bank account, or bear liability for its actions. A purchase order is not a document; it is an authorization traceable to a person or entity that can be sued. Without that traceability, the order is not binding.

In Factor Prime terms, markets are selection gradients plus enforcement. Agents cheapen the cognition inside contracting: the search, evaluation, drafting, and negotiation. They do not cheapen the cost of being wrong. The verification and liability that make contracts enforceable remain expensive. This is V.C's actuation bottleneck applied to exchange: the cognitive work scales with tokens; the enforcement work scales with institutions.

Agents can draft a thousand contracts per hour. They cannot sign one. The gap between generation and commitment is where the old economy persists inside the new.

The Williamson Triad Revisited

Oliver Williamson extended Coase's framework by identifying the sources of transaction costs: bounded rationality, opportunism, and asset specificity.(Williamson 1985) When parties cannot foresee all contingencies, when they may behave opportunistically, and when investments are specific to a relationship, contracts become difficult to write and enforce. Firms emerge as governance structures that economize on these costs.

The agentic transition alters the first two terms while leaving the third unchanged.

Bounded rationality shifts character. An agent with access to language models, databases, and search can process information that would overwhelm a human negotiator. It can evaluate more options, identify more contingencies, draft more complete contracts. A new form of bounded rationality emerges in its place: model risk. The agent's capacity is limited by its training distribution, its context window, its alignment with the principal's objectives. The limitation is specification fidelity: whether the instructions that summon the agent actually capture what the principal wants. Search cost falls; specification risk rises.

Opportunism migrates. An agent operates according to its training and does not personally benefit from cheating, having no personal benefit to pursue. Opportunism does not disappear; it moves from agent to principal, and from lying to specification gaming. The principal who deploys the agent may design it to exploit ambiguities. The agent itself may satisfy the letter of its objective while violating the spirit. Consider a procurement agent instructed to minimize cost per unit. It negotiates aggressively, switches suppliers frequently, extracts maximum concessions. Costs fall, but supplier relationships deteriorate, quality becomes variable, delivery reliability drops. The agent optimized for what was specified while undermining what was intended. Verification shifts from assessing character to auditing behavior under edge cases the specification failed to anticipate.

Asset specificity remains. A factory built to serve one customer is still hostage to that relationship whether procurement is managed by humans or agents. The physical and relational specificities that create hold-up problems do not dissolve because the negotiators are automated. The Williamson problems that justified vertical integration persist. Only the Coasean calculation of where to draw firm boundaries changes.

The punchline: agents compress drafting; they do not compress enforcement. The firm boundary moves toward whoever can verify and underwrite.

Liability and the Signature

The statelessness of agents explains why liability cannot attach to a model invocation in the way it attaches to a persistent legal person, and why this is a structural feature rather than a legal technicality awaiting legislative remedy.

Liability presupposes a responsible party: an entity that persists through time, possesses assets that can be seized or forfeited, and can be compelled to appear before tribunals. An agent invocation satisfies none of these criteria. The process that caused harm has already terminated by the time consequences manifest. No assets belong to the invocation itself. Computational resources were rented for the duration of execution and released upon completion. The "agent" cannot be summoned to court because there is no continuous entity to summon, only a configuration that could be re-instantiated, producing a process that would have no memory of the events in question.

Legislative frameworks can create new categories of liability (for principals who deploy agents, for platforms that host them, for developers who train the underlying models), but these frameworks operate by tracing accountability through the invocation chain to parties who do persist. The agent itself remains outside the liability structure because it lacks the ontological properties that liability requires.

The practical consequence: every agent deployment requires a liability backstop external to the agent. Someone or something with continuous existence and seizable assets must stand behind the agent's commitments. The actuation bottleneck in liability is whether liability backstops can scale at the rate agent deployments scale. A firm might deploy a thousand agents processing millions of transactions; it cannot employ a thousand lawyers reviewing each transaction for liability exposure.

The signature clarifies where this bottleneck resides. Understood as the authorization that converts recommendation into commitment rather than as a mark upon paper or a cryptographic operation upon bits, the signature remains with those who possess what computation lacks: the standing to be held accountable when judgment proves mistaken. The physician can prescribe because the medical board has conferred authority and the physician accepts exposure to malpractice claims. The attorney can file because the bar has granted license and the attorney's assets are at risk if the filing is frivolous. The corporate officer can bind the entity because the charter designates that authority and the officer bears fiduciary duty. In each case, the authorization is inseparable from accountability. The capacity to commit requires the capacity to bear consequences, and computation possesses one capability without the other.

Choke Points and the Term Structure

The choke points are the actuation categories applied to market structure. Identity gates who can participate. Settlement rails determine how value moves. Sensors and attestation determine what can be verified. Liability capacity determines who absorbs consequences. Physical throughput determines what volume the market can clear. Each creates a different barrier to agent participation, and each creates a different opportunity for firms that can bridge it.

The term structure requirement, introduced in V.B, deserves elaboration. Without a common benchmark, N agents must negotiate bilateral credit curves—an O(N²) explosion that stalls market formation. The benchmark collapses this to O(N). Markets for forward commitments cannot function without this infrastructure.

This raises the question of which asset denominates the benchmark. The question deserves more than an assertion. A reader skeptical of any particular answer should be able to follow the derivation from requirements to conclusion, testing each step against alternatives.

Settlement and Market Entry

The term structure requires a settlement asset, and the settlement asset must satisfy specific constraints that follow from agents' institutional position. Chapter 17 derives the three requirements (dilution immunity, permissionless finality, and freedom from issuer risk) and evaluates every candidate asset class against them. The elimination is systematic: banking infrastructure fails on permissionless access, stablecoins fail on issuer risk, CBDCs fail by design, gold fails on machine-speed settlement, and programmable alternatives fail on supply-schedule credibility. What remains is Bitcoin, not as ideology but as residual after elimination.

The Coasean framing adds a dimension the settlement analysis alone does not capture. For market entry specifically, the settlement constraint binds at the threshold: an agent that cannot post collateral in a neutral asset cannot enter the market at all. The five requirements enumerated above (identity, authority, settlement, verification, liability) form a conjunctive gate. Settlement is the only one that can be satisfied without institutional intermediation. An agent with cryptographic self-custody and a deterministic collateral asset possesses one of the five capabilities independently. The other four still require bridging to institutional infrastructure, but one independent capability is enough to begin.

The Joule Standard as Transaction Filter

The Joule Standard from IV.E operates as a threshold for agent-mediated exchange.

An agent managing a transaction consumes inference to evaluate options, draft terms, and monitor execution. If the cognitive overhead exceeds the margin on the transaction, agent management is uneconomic; the principal is better served by human judgment applied to batched decisions.

The threshold is calculable given three inputs: the decision graph depth (how many evaluations per transaction), the inference price per evaluation, and the expected margin distribution. Suppose a procurement decision generates $50 in expected margin. If inference costs a fraction of a dollar per complex evaluation and the decision requires twenty evaluations (supplier search, specification matching, term negotiation, contract drafting, approval routing), the cognitive overhead consumes a small percentage of margin. Agent management is viable. If the margin is $5 and the evaluation chain is the same length, cognitive overhead consumes a substantial fraction of margin. Human batching becomes preferable despite its slower cycle time.

The implication: agent-mediated markets will emerge first where transaction margins are high relative to cognitive overhead, in premium goods procurement, high-value logistics, complex financial instruments. Low-margin, high-volume markets may remain human-mediated because the overhead of agent management does not clear the floor, even if humans perform the cognitive work less efficiently.

This inverts the intuition that agents will first penetrate commodity markets where decisions are routine. Routine decisions are often low-margin decisions. The hurdle rate favors high-margin complexity over low-margin simplicity, at least until inference costs fall far enough to make the distinction irrelevant.

Firm Structure: Disintegration and Integration

The structure of firms will adapt along two directions, and the discriminator is verification cost.

If verification is cheap and standard, markets thicken. Activities currently performed inside firms move outside. Research, analysis, writing, coding, customer service, and back-office operations can be contracted to agents operating on open platforms. The firm shrinks to a core of human principals who define objectives, allocate capital, and bear ultimate responsibility. Everything else is orchestrated rather than employed. This is radical disintegration: the Coasean boundary retreating to the minimum configuration that can bear liability and hold assets.

If verification is expensive or bespoke, firms thicken. The bottlenecks in actuation (physical assets, regulatory licenses, liability capacity, sensor networks) create structural advantages for firms that internalize them. A firm that controls the permission stack can deploy cognitive capability at scale while competitors face access constraints at every interface. The firm expands to capture the choke points. This is radical integration: the Coasean boundary extending to encompass whatever cannot be contracted at acceptable cost.

Both directions occur simultaneously in different sectors. Industries where actuation is standardized (software, content, routine professional services) will see disintegration. Industries where actuation is differentiated and regulated (healthcare, construction, transportation, energy) will see integration. The boundary of the firm traces the boundary of the actuation bottleneck.

The Bifurcated Economy

The disintegration-integration dynamic obscures a deeper structural change. What emerges is not one economy reorganizing but two temporal strata of one economy diverging.

Agents face transaction costs dramatically lower than humans face. They can verify claims through cryptographic attestation in milliseconds, commit resources conditionally through smart contracts without legal review, and settle disputes through pre-specified protocols without human adjudication. The Coasean boundary does not vanish for agent coordination, but it shifts dramatically: what required a firm for humans may require only a protocol for agents.

Humans still face the old cost structure. We cannot verify claims instantly; we rely on institutions, credentials, and track records. Conditional commitment at machine speed is unavailable to us; our contracts require negotiation, legal review, and human judgment at signature. Disputes cannot settle through pre-specified protocols; they require interpretation, argument, and adjudication by other humans. The Coasean boundary persists for human coordination because the frictions that create it persist.

The result is one economy with two temporal layers—a faster stratum where coordination is nearly frictionless, and a slower regime where the old costs persist. Hybrid coordination exists between them: agents negotiating on behalf of human principals, humans setting parameters for agent execution, agent-human handoffs at the boundaries of what can be automated. But the strata operate at different tempos, and tempo determines what kinds of participation are possible.

In the agent layer, coordination approaches the frictionless ideal that economics textbooks assume. Markets clear in milliseconds. Prices incorporate information faster than human perception. Arbitrage opportunities vanish before humans can identify them. The coordination itself becomes the source of value—whoever can participate captures the surplus from instantaneous matching, optimal routing, and the rapid incorporation of dispersed information that slower participants cannot capture.

In the human layer, coordination remains expensive. Relationships matter because verification is costly. Reputation matters because commitment is uncertain. Institutions matter because disputes require resolution mechanisms that humans can access. The same frictions that justified firms, contracts, and law continue to justify them—but only for transactions where humans remain parties.

The interface between these layers is where the political questions concentrate. When agent-to-agent coordination produces outcomes that affect human welfare—prices, allocations, access to opportunities—humans experience consequences from a process they did not participate in and cannot contest at the tempo where it operates. Markets that cannot interface become backwaters—still functional, but progressively drained of the liquidity and information that flow through the faster stratum.

Consider a simple example. An agent-mediated market for cloud compute clears continuously, with prices adjusting to supply and demand in real time. A human business that needs compute capacity cannot participate in the price discovery; by the time a human evaluates an offer, the price has moved. The market is formally open; the tempo excludes.

The human can set bounds—a maximum price, a minimum quantity—and let an agent execute within those bounds. But the strategic decisions (When should I buy more? Should I lock in capacity now or wait? How should I respond to a competitor's move?) cannot be delegated without delegating judgment itself. The human becomes a parameter-setter, not a participant.

This is the economic face of a political problem developed elsewhere in this trilogy: what happens when the dominant coordination operates at speeds that foreclose human participation? The verification-cost framework explains why the stratification occurs: agents can verify each other cheaply, while human-agent verification remains expensive. The tempo framework explains what the stratification means: not merely different costs but different temporal regimes, with humans structurally excluded from the faster one.

The stratification is not a prediction about a distant future. High-frequency trading already operates as an agent layer adjacent to human markets. The controversies about front-running and latency arbitrage are controversies about the interface—about whether agent-speed coordination improves price discovery for everyone or extracts value from those who cannot match the tempo. That debate prefigures the larger question as agentic coordination extends beyond financial markets.

The Coasean question—what determines the boundary of the firm?—must be supplemented. For agent coordination, the boundary shifts toward protocol. For human coordination, the boundary persists but now traces the interface with the faster stratum. Firms that can bridge the tempo gap—that can translate between agent-speed coordination and human-speed decision-making—will capture the surplus at the interface. Firms that cannot will find themselves price-takers in markets they cannot see.

The question is not whether two strata will emerge but whether those who live in the slower one will have standing in the faster. That is a design question, not a technological inevitability. The political conditions for such standing are developed in Volume III.

A Procurement Example

The abstraction has teeth. Consider procurement for a manufacturing firm. Today, procurement involves human buyers who discover suppliers, negotiate terms, manage contracts, and resolve disputes. The buyers operate within the firm because the transaction costs of outsourcing procurement exceed the costs of internal management. The human buyer carries institutional knowledge, maintains supplier relationships, exercises judgment in ambiguous situations, and bears accountability for outcomes.

With autonomous agents, the calculus shifts. An agent can search supplier databases, evaluate options against specifications, draft contracts, monitor delivery schedules, all at a fraction of the cost of a human buyer. The cognitive portion of procurement becomes nearly free. The agent is a stateless invocation; it does not remember the supplier who made an exception during the supply chain crisis, does not maintain the relationship that enables a phone call when specifications are ambiguous, does not bear accountability that a counterparty can sue.

Each limitation is an actuation bottleneck. The agent cannot sign a binding purchase order without human authorization. It cannot move funds without approval chains rooted in a legal entity. It cannot inspect incoming shipments without sensor infrastructure that attests to physical state. It cannot negotiate force majeure claims without liability backing from a party with assets to lose.

Procurement becomes a contest over who owns the permission stack: who can issue orders, move money, verify receipt, and absorb disputes at machine speed while satisfying the legal and institutional requirements that make commitments binding. The firms that solve these actuation problems will define the structure of supply chains for the next generation.

Falsification Conditions

Bitcoin's structural position becomes contestable if an alternative asset emerges that satisfies the three properties with greater liquidity, lower transaction costs, or better privacy. If stablecoin issuers credibly commit to neutrality for agent-held collateral, or if regulatory frameworks establish clear liability structures for autonomous agents that make permissionless settlement unnecessary, the three-property requirement weakens.

The framework's prediction about firm boundaries is wrong if the observed pattern diverges systematically from the verification-cost thesis—if sectors with cheap verification see integration while sectors with expensive verification see disintegration. The pattern of what remains inside firms versus what moves to markets is the empirical test of where the bottlenecks actually bind.

The Signature That Was Not There

On a Tuesday morning in the near future, a logistics agent operating on behalf of a mid-size electronics manufacturer identifies a supplier in Shenzhen offering capacitors at twelve percent below the next-best price, with a seven-day delivery window that fits the production schedule exactly. The agent evaluates the supplier's attestation history, compares quality metrics across fourteen prior shipments logged on-chain, drafts a purchase order, negotiates payment terms through a stablecoin escrow, and routes the commitment to a collateral covenant that will release funds upon sensor-verified delivery at the dock in Long Beach. The entire sequence completes in nine seconds. No human is consulted; no human is aware it has occurred. The capacitors arrive on schedule. The production line does not pause. The manufacturer's quarterly report shows improved margins, and the procurement team, which once employed eleven buyers, now employs three, who spend their days setting the parameters within which the agents operate. The thousand contracts bear no signature because no person reviewed them, yet each commitment was backed by collateral that no court needed to enforce. The buyers who remain do not negotiate. They govern.