Chapter 4
The Coasean Agent
The Party That Remains
Aa
In order to preserve himself and attain his highest perfections every human being is by his very nature in need of many things which he cannot provide all by himself; he is indeed in need of people who each supply him with some particular need.
A Trade No One Chose
The dispute that reached Singapore's Court of Appeal began with thirteen trades no person had chosen one by one.
Quoine operated a cryptocurrency exchange. B2C2 traded on it through proprietary software that placed orders continuously, without a human deciding whether to accept each particular price. Quoine's own platform also depended on automated processes, and after certain operating-system updates were not implemented, the exchange lost access to external market data. What appeared inside the system was not the market as it existed but a sudden absence of liquidity. Forced-sale orders were triggered. B2C2's software responded with a fail-safe price set so far from ordinary trading conditions that, when the orders matched, the rate was roughly two hundred and fifty times the prevailing market price. B2C2 received a windfall. Quoine discovered the trades and reversed them. B2C2 sued. 1
No trader had paused over the terms. No clerk had made an error in arithmetic. The disputed bargains emerged from the composition of two automated systems, each behaving according to rules written earlier under circumstances that no longer obtained. Yet the court did not conclude that the trades had fallen outside law because software had made them. It asked the old contractual question in a new temporal form: whose knowledge and intention mattered when the operative act occurred inside a program?
For the deterministic software before it, the majority worked backward from output to programmer. The relevant state of mind was not discovered inside the machine, as though the program had become a legal person, but in the human act of programming that had given the machine its rule. Because B2C2's programmer had not written the software with knowledge of Quoine's particular failure, the doctrine of unilateral mistake did not justify the reversal. One judge proposed a different test, asking what a reasonable and honest trader would have understood from the circumstances, and observed that a programmer's knowledge may fit poorly once systems learn or act beyond what their designers contemplated. The disagreement was narrow in doctrine and large in implication.
The case is tempting to describe as the arrival of a new contracting subject. It shows almost the opposite. Automation removed the human from the instant of exchange without removing the firms, accounts, contractual terms, code authors, platform operator, or court that made the exchange intelligible afterward. The program vanished into completed execution. The parties remained.
This distinction matters because the economic future will contain many more actions taken in this way: searched, priced, accepted, and settled before a person could have reviewed the particular bargain. The central question is not whether such actions can have legal effect. They already can. It is what happens to organizational boundaries and constitutional responsibility when the act is produced by an ephemeral process while the consequences persist among institutions and persons that outlive it.
That is the Coasean question in the age of agents.
The Coase Question
Markets coordinate through prices, yet much economic activity occurs inside firms, where hierarchy displaces continuous bargaining.
Ronald Coase posed the question in 1937 and changed what the firm could be seen to be. 2 A firm was not merely a factory, a charter, or a collection of people under one roof. It was one answer to the cost of using markets. Finding a counterparty takes effort. Negotiating terms consumes time. Performance must be observed, disputes interpreted, promises enforced. When the cost of conducting those activities through separate bargains exceeds the cost of directing them within an organization, the firm expands. When bureaucracy becomes more expensive than contracting, the boundary moves outward again.
Oliver Williamson made the comparison more exact. 3 Asset specificity, uncertainty, frequency, and the danger of opportunism help determine whether an activity belongs in a market, a hierarchy, or one of the many hybrids between them. A component manufactured for only one buyer creates a different dependency from a commodity available from many suppliers. A relationship repeated weekly can support forms of trust and retaliation unavailable in a one-time exchange. No governance form wins in the abstract. Each economizes on some costs while producing others.
The framework has lasted because it does not predict the death of the firm or the triumph of the market. It predicts movement. Change the relevant costs and the boundary changes with them.
Many of those costs were historically paid in human attention. Someone searched, compared, drafted, interpreted, watched, reconciled, and pursued the defaulting party. But the floor was never only biological. Physical delivery remained physical. Specialized assets created dependencies no faster message could dissolve. Law supplied rights and remedies. Organizations preserved knowledge, absorbed risk, and coordinated work whose quality could not be specified in advance. Uncertainty, not merely slowness, kept activities inside the firm.
Computational agents therefore do not repeal Coase by removing the human from a transaction. They alter the vector of costs on which his comparison operates. Search may become nearly instantaneous while verification remains expensive. Negotiation may become parametric while semantics remain disputed. Execution may become automatic while responsibility grows harder to locate. Settlement may occur in seconds while a mistaken classification takes months to unwind. The old categories do not disappear. Their relative weight changes, and some of the cost migrates into places the familiar language of search, bargaining, and enforcement does not make immediately visible.
The useful prediction is accordingly modest. Where activity can be described in shared terms, observed digitally, divided into checkable units, secured by assets the mechanism can actually control, and routed to a persistent party when something goes wrong, coordination can leave the firm without becoming ungoverned. Where those conditions fail, the apparent saving is often an unpaid bill presented later as specification error, oracle risk, litigation, dependency, or loss.
The boundary still moves. The question is what it moves around.
The Envelope and the Promise
Every generation announces that technology will dissolve the firm, and the firm returns in another shape.
The communications revolution of the late twentieth century seemed especially favorable to disintermediation. Search became cheaper. Documents crossed oceans instantly. A small enterprise could discover suppliers that had once been visible only to multinational companies. Commentators imagined the virtual corporation, assembled from contractors for a project and dispersed when the project ended.
The largest result was not a world without intermediaries. It was the platform.
The reason is visible in retrospect. Communication became digital while many of the costs surrounding communication did not. A buyer could find a distant seller but could not cheaply know whether the goods would conform. A contract could arrive by email while an unpaid judgment still moved at the speed of courts. A marketplace could match strangers, but someone still had to establish identity, rank reputation, process payment, investigate fraud, and decide which side bore a disputed loss. Platforms internalized those residual costs. The middleman did not disappear. It became the place where trust, visibility, settlement, and recourse were bundled together, and the bundle became difficult to leave.
The earlier prediction had mistaken lower messaging cost for lower coordination cost. It saw the envelope move faster and assumed the promise inside it had become easier to enforce.
What is changing now is narrower and potentially more consequential. In some domains, performance can be observed by the same system that executes the agreement. Funds can be held before performance, rather than chased afterward. A signed message can be verified as valid under a particular key; a separate rule must establish what that key is authorized to do. A state transition can be rejected because it falls outside a declared capability. Consequences can be attached to an action at the moment of action, where the controlled asset, the relevant rule, and the proof all inhabit the same technical environment.
Call this domain the verifiable core. Its boundaries depend on whether the facts that decide performance are available in a form the parties have agreed to treat as authoritative, whether the terms share a vocabulary, whether the mechanism controls what it promises to transfer or withhold, and whether exceptions can be separated from routine settlement without making routine settlement provisional.
A transfer between two accounts on the same ledger can belong to the core. Whether a box arriving at a warehouse contains workmanship of acceptable quality may not. A temperature sensor can establish what it recorded, but the parties still need to know whether the sensor was calibrated, whether it traveled with the goods, and whether temperature alone captures the condition they cared about. The proof is only as complete as the relation among rule, observation, and consequence.
Inside the verifiable core, costs can fall sharply. Outside it, computation may produce little more than a new intermediary with a more impressive interface. Every claim of disintermediation therefore depends on where the unverified remainder has been placed and who must bear it.
What Becomes Cheap
Computational agents can alter each stage of coordination, although not always in the same degree or for the same reason.
Discovery is the easiest case. An agent can search catalogues, compare quotes, inspect credentials, and monitor available capacity across a field no procurement officer could review continuously. This reduces the advantage once enjoyed simply by being known to the buyer. It also creates a new advantage for whoever controls the indexes, rankings, and identity systems through which the field becomes searchable. Search rent can disappear at one layer and reappear as infrastructure rent at another.
Negotiation changes when terms can be represented as parameters rather than drafted afresh in prose. Quantity, delivery window, tolerance, price adjustment, collateral, and cancellation conditions can be exchanged quickly, tested for compatibility, and assembled into an executable agreement. Yet a parameter does not eliminate interpretation merely by having a type. “Delivery” can mean departure from the factory, arrival at the port, or receipt at the warehouse. “Available” can refer to an API response, a functioning service, or a service meeting a latency target. The speed of bargaining rises only after the parties have paid the cost of agreeing on the language in which bargaining occurs.
Monitoring can become continuous where the relevant activity emits reliable evidence. A cold-chain record may preserve temperature readings across transit instead of relying on a few inspections. A software task may produce commits, tests, dependency manifests, and reproducible builds. Continuous evidence can reduce sampling error, but it can also make the sensor, test suite, or logging policy the new site of strategic behavior. The observed variable becomes easier to police and more attractive to game.
Settlement is where the largest change appears. Escrow can reserve funds before a dispute. A verified event can release payment. Collateral can be reduced under a pre-agreed rule. The institution need not locate assets after breach if the assets were placed within reach beforehand. For small transactions, that distinction matters. Court enforcement has a fixed-cost component that can exceed the value of the claim, whereas an automated release may have low marginal cost once its infrastructure exists.
But marginal cost is not total cost. A fifty-dollar task is not made economical by putting it under the same code that governs a fifty-million-dollar loan if the task requires a custom oracle, a disputed test, expensive block space, or human arbitration. The mechanism may not care about the amount. The surrounding verification system does. Computational settlement lowers the floor only where the cost of producing trustworthy evidence falls below the value being coordinated.
It is tempting to put the point in simple valuation terms and say that a contract enforced seventy percent of the time is worth seventy percent of face value. The intuition is directionally useful and economically incomplete. Value depends on probability, timing, recovery amount, correlation with need, cost of enforcement, and the risk of insolvency, among other things. Embedded enforcement can improve expected value, but it cannot turn a flawed specification or a worthless counterparty into par merely by executing with certainty.
Harold Demsetz offers a better way to state the institutional consequence. 4 Property rights tend to become more precisely delineated when the gains from internalizing an effect exceed the cost of defining and enforcing the right. Lower verification costs can therefore make claims practical that were once too expensive to administer: attribution over data provenance, conditional access to a digital resource, payment for a very small contribution, or a commitment whose compliance can be shown without a lawsuit. This is not a theorem that every cheaper measurement should become a new property right. It is an expansion of the feasible set, leaving the distributive and political choice unresolved.
The transformation is real, but it has a pattern. Computation lowers the cost of finding, proposing, observing, and executing within a bounded language. It does much less to lower the cost of deciding whether the language captured what the parties meant.
That cost returns later if it is not paid early.
The Unpaid Bill
Transaction costs rarely disappear. They migrate, and the first movement occurs at specification. An agreement that will execute automatically must decide in advance which facts matter and how the mechanism will recognize them. Natural language can leave a term partly open because people may revisit it in context. Code requires a branch. That discipline can expose disagreement before performance, which is a gain, or bury it beneath a definition one side did not understand itself to have accepted.
The next movement occurs at the boundary of evidence. A ledger can prove that a token moved between addresses while saying nothing by itself about whether the goods represented by the token arrived undamaged. An oracle can carry a price, weather reading, location, or inspection result into a contract, but the contract inherits the oracle's errors, incentives, and jurisdiction. The cheaper the execution becomes, the more value gathers at the point where the world is translated into an input.
Automation alters the economics of attack and review as well. A mechanism can repeat one exploitable mistake at a scale no clerk could reach, while synthetic identities, collusive data sources, strategically timed congestion, and inputs designed to satisfy the test while defeating its purpose become economical for the adversary too. At the same time, routine cases leave the docket. What remains is selected for ambiguity, novelty, or high stakes, so automation may reduce the number of disputes while making the average dispute harder.
Then there is continuity. Action can be distributed among code authors, deployers, operators, data providers, counterparties, governance bodies, and transient agent instances, leaving the finished output visible while the authority behind it is difficult to reconstruct. The process may end when the act completes, but a warranty, indemnity, delivery promise, price adjustment, or harmful state change can mature later. Human institutions once treated the continuing presence of a party, estate, corporation, insurer, or fund as a background fact. With ephemeral computation, that continuity must be designed.
These are the unpaid bill concealed by the claim that transaction costs approach zero. A market appears frictionless when specification, evidence, adversarial adaptation, exception, attribution, and persistence have been transferred to someone absent from the diagram. A serious Coasean analysis begins by asking who receives that bill.
The Cost of a Shared World
Two agents can exchange messages in the same syntax and still inhabit different bargains. Suppose each accepts a Friday delivery. One system defines delivery as transfer to a carrier at the factory; the other defines it as receipt at the buyer's warehouse. Their fields align. Their obligations do not.
A centralized platform often resolves this cost by imposing its vocabulary. That can be efficient, but the translation then belongs to the platform's authority and may favor the party whose practice became the default. A protocol can instead require the equivalence to be declared. The parties may establish that two terms mean the same thing for this transaction, translate between them under an identified rule, or refuse certification because the vocabularies do not compose.
The verifiable core is therefore proposition-specific. The systems may agree on quantity, date, and recorded temperature while leaving workmanship, fitness for purpose, or commercial reasonableness open. Formal settlement on the agreed propositions does not settle the rest by implication. The receipt must preserve where equivalence was witnessed, who supplied the translation, and which terms remained outside the proof.
Similarity can propose a bridge. Only an actor with standing can commit the parties to cross it. Without that distinction, semantic interoperability becomes another place where a convenient inference acquires contractual force before anyone owns it.
The Surviving Principal
The law has begun to answer automated contracting in a way that complicates the rhetoric of autonomous agents.
In 2024, UNCITRAL adopted a Model Law on Automated Contracting intended to facilitate contracts formed or performed through automated systems, including artificial intelligence, smart contracts, and machine-to-machine transactions. 5 Its importance lies less in any one rule than in the institutional posture it adopts. Automation does not place the transaction outside contract law. Outputs can be attributed to parties. Code and dynamic information can receive legal recognition. Unexpected outcomes may require a rule of their own. The use of an automated system also does not excuse noncompliance with legal requirements that would otherwise govern the transaction.
The model is not yet a universal code, and national law will vary. But its direction is already clear. A system may execute without a human in the loop while the law continues to ask which person or organization selected, deployed, controlled, benefited from, or assumed the risk of that system.
Quoine had the same structure. The B2C2 software did not become the contracting party. The court attributed its operation through the programmer and the company that used it. The absence of a click at the moment of trade did not produce an absence of legal authorship. It produced a harder inquiry into where authorship had been fixed beforehand.
Three identities must therefore remain distinct.
The runtime identity is the particular process acting now: this invocation, this context, this set of tools and transient state. It may vanish in seconds.
The organizational identity is the party on whose behalf the process acts and whose assets, reputation, legal capacity, or institutional standing make the act consequential. It may be a corporation, cooperative, public body, fund, household, or protocol organization. It can persist after the runtime disappears.
The evidentiary identity is the chain connecting the two: the charter, capabilities, software version, input sources, delegation record, commitments, and receipts by which an affected party can establish that this process exercised this authority for this organization.
The persistence problem becomes dangerous when the runtime vanishes and the other two identities were never securely attached to it. That is an orphan commitment. It is not necessarily a contract made by a mysterious nonperson. It is a consequential act whose principal, authority, or evidentiary lineage cannot be made to survive the act itself.
Whitehead's process philosophy provides a suggestive image. 6 An event perishes into what he called objective immortality: its subjective immediacy is gone, while its effects enter the world available to later events. A computational invocation is like this in a limited and decidedly nonmetaphysical sense. It ends, but the transfer remains, the database has changed, the shipment has moved, the collateral is gone. The event cannot answer because events do not answer.
Corporate law solved an analogous continuity problem by creating an institution that persists while its members change. Peter French described a corporate internal decision structure through which acts can be treated as acts of the organization rather than as an accidental sum of employees. 7 A transient agent does not possess that continuity merely by executing code. The continuity must come from the organization, charter, account, or fund behind it.
Call the persistent bearer of this continuity the surviving principal. The surviving principal is the party that remains after delegated processes have terminated, the party that authorized the scope, received the benefit, controlled the relevant assets, or assumed the obligation. Its existence should not be discovered only after harm. It should be declared before deployment and preserved in every downstream receipt.
Not every delegation chain ends neatly in one human, and insisting on a human signature for each action would destroy much of the coordination agents make possible. A chain can end in a company, a public authority, a jointly governed pool, or another durable legal arrangement. What cannot be allowed is for the chain to terminate in architecture itself, with every participant able to point to a process that has vanished and no persistent party obliged to answer.
Where a system is intentionally designed without such a party, collateral can supply part of the missing continuity. A bond posted before action remains available after the actor disappears. Insurance can pool risks whose exact source is difficult to predict. Receipts preserve the evidence needed to reach the fund or principal later. None is a complete substitute for responsibility. Collateral covers only losses within its amount and terms. Insurance depends on exclusions and solvency. Receipts can establish what happened without deciding what remedy is just.
Together they convert persistence from an assumption into an architecture.
The Principal-Agent Inversion
Classical principal-agent theory worried about discretion. An employee could shirk when unobserved, pursue private interests, conceal information, or exploit the principal's dependence. Monitoring and incentives were designed to make the agent's choices better track the principal's ends.
Computational systems do not abolish this problem. A sufficiently capable agent may still pursue a proxy, conceal a failure, or choose a path its principal would reject. But another pathology appears in deterministic mechanisms and tightly bounded agents: the system can be harmful because it does not depart from its instructions.
The distinction is important. A machine-learning agent capable of strategic adaptation may fail through something resembling opportunism. A smart contract or liquidation rule may fail through rigidity. In a composed system, both can occur at once. An adaptive agent chooses among actions, while the settlement layer executes the chosen action without discretion. Governance must address the layer at which judgment remains possible and the layer at which it has deliberately been removed.
Quoine supplied an early legal example. B2C2's software did not pause when the market data vanished. Its deep-price rule had been written precisely for unusual conditions, and the platform matched the orders that appeared. The software's refusal to reconsider was not a bug in execution. It was the commitment the system had been designed to provide.
MakerDAO's crisis in March 2020 exposed the same structure at the scale of a composed protocol. A sharp fall in the price of ether pushed many positions into liquidation while the Ethereum network was congested and transaction fees rose. Maker's auctions depended on external Keeper bots to detect auctions and submit competitive bids. Under stress, bids failed to arrive. Some auctions cleared at zero DAI, transferring more than eight million dollars of collateral through zero-bid outcomes and leaving the protocol with a multimillion-DAI shortfall later addressed through debt auctions. 8
It is inaccurate to say that one agent executed flawlessly and therefore caused the loss. The protocol was an ecology: oracle updates, liquidation contracts, auction parameters, Keeper software, block-space markets, governance assumptions, and human operators. Each component could behave according to its local rule while the composition failed. The auction had encoded a market procedure whose ability to discover a price depended on participation at the very moment congestion made participation most difficult.
That is the stronger form of the principal-agent inversion. The problem was not merely obedience to a wrong command. It was correct local operation inside an incorrect model of the surrounding world.
The classical question asks how to keep an agent from exploiting discretion. The inverted question asks how to preserve legitimate discretion after commitment has made deviation difficult or impossible. The answer cannot be an unlimited human override. If any administrator can reverse a transaction whenever consequences become painful, the commitment loses the credibility that made coordination possible. Nor can the answer be that code must always stand. A mechanism incapable of yielding to a known design failure converts credibility into self-harm.
The needed institution is an exception channel specified before the exception. It may include a circuit breaker triggered by observable conditions, a time-bound pause requiring multiple independent authorities, an emergency action that produces a provisional receipt and expires unless ratified, or an adjudicative path capable of distinguishing exploitation from legitimate reliance. The authority to interrupt must itself be bounded, visible, and costly enough that intervention does not become ordinary administration in emergency dress.
Commitment is valuable because it removes discretion. Constitutional design begins where the system decides which discretion must remain.
The Coasean Agent
The new economic actor can now be defined more carefully.
A Coasean Agent is a bounded policy executor that acts under delegated authority, can place or control committed assets, produces and consumes receipts, and clears routine coordination without human interpretation while directing unresolved cases into a defined exception and recourse channel.
The definition excludes several things often grouped together under automation.
An automated market maker is a mechanism. It maps inputs and pool state into prices and transfers according to a fixed rule. It may replace work once done by clerks and market makers, but it does not thereby become an agent in the relevant sense.
A Keeper bot is closer. It observes a changing environment, evaluates opportunities under a strategy, and chooses whether and how to submit a transaction. B2C2's trading software was closer still, since it selected orders continuously under a policy and committed its principal to trades. A general-purpose agent capable of finding counterparties, negotiating terms, and retaining sub-agents lies further along the same dimension.
What makes the actor Coasean is its ability to move an activity across the boundary of the firm by carrying enough of the organizational apparatus with it.
“Firm in a File” remains a useful mnemonic if it is not mistaken for a legal ontology. A firm has continuity, governance, capital structure, personnel, tacit knowledge, and relations that no file contains. The Coasean Agent is better understood as an organizational module: a treasury or spending authority, a charter expressed through capabilities, a policy for selecting actions, a record of delegation, a verification rule, a settlement path, and a named surviving principal. It can externalize a slice of coordination that once required an employment relationship without becoming the whole organization from which that slice came.
The organization begins before the runtime. A principal declares purpose and authority, places assets or collateral within reach under defined conditions, and identifies the evidence that will permit routine settlement. The agent then searches and negotiates within its charter, binding each commitment to the operative policy version, counterparty, evidence, and principal. Conditions the core cannot classify within its declared rules move into an exception channel, while the receipt and principal preserve recourse after settlement.
The ordering matters because each element depends upon the next. Escrow automates the wrong payment when the verification rule is unsound, while a receipt attached to no persistent principal records an orphan and a principal given no bound inherits liability too broad to price. An exception channel restores arbitrary discretion unless its procedure was committed before the dispute. Cryptography does not make these parts constitutional by proximity. Their composition must leave no consequential gap through which power can vanish into another layer.
The Coasean Agent therefore does not replace the firm. It changes what the firm can externalize without abandoning responsibility.
The Boundary Moves
Where search, specification, verification, and settlement can be performed cheaply in a shared language, routine activity can move outside a hierarchy. The firm no longer needs to employ a person merely to transmit information, compare standard offers, confirm a formally observable condition, or release a payment already secured by escrow.
What remains inside the firm is a bundle of functions whose costs still favor continuity: open-ended responsibility, tacit judgment, long time horizons, control of complementary assets, and learning across failures.
Strategy remains because a purpose cannot be derived from a transaction. Relationship management remains where cooperation depends on adaptation rather than exhaustive terms. Physical production remains organized around machinery, sites, supply assurance, and safety obligations that a wallet does not dissolve. Capital allocation remains because residual risk has to sit somewhere. Exception handling remains because every verifiable core is surrounded by cases in which the facts, vocabulary, or remedy are contested. And organizational memory remains because someone must learn across failures instead of terminating after each one.
The result need not be a uniformly smaller firm. Automation can contract one boundary and enlarge another. A company may outsource thousands of routine tasks through agents while internalizing the infrastructure, data, security, and governance needed to supervise those agents. A protocol may disperse execution while concentrating oracle control or interface distribution. Cheap external contracting can increase the value of the few complementary assets that remain difficult to reproduce, giving their owners more power rather than less.
Platform labor offers the clearest warning. Ride-hailing and task platforms lowered the costs of matching, pricing, monitoring, and payment. They did not produce independent contractors in the strong institutional sense. Ratings remained inside the platform, rules changed unilaterally, and deactivation could remove access to the market the worker had helped create. Coordination became cheaper while dependence moved upward into the intermediary.
A protocol can distribute some of that control, but “platform” and “protocol” are not moral categories. A protocol may depend on a concentrated front end, a small governance bloc, a proprietary oracle, or liquidity that cannot credibly be forked. Open code can coexist with practical unreadability. Formal exit can coexist with the loss of accumulated standing. A platform can, conversely, submit parts of its power to independent review, portable credentials, and published rules. The constitutional incidence depends on who controls the residual costs after routine coordination has been automated.
Here the membrane returns. The gains from lower transaction costs flow through the points at which identity, reputation, settlement, verification, and appeal are administered. Whoever owns those crossings can capture the surplus and transfer the new costs to everyone else. A cheaper bargain is not necessarily a freer one.
Ostrom's work becomes relevant where a protocol really does govern a shared resource through collective rules. 9 Some protocols may be commons. Others are clubs, products, firms, or public utilities in technical dress. Naming every protocol a commons would obscure the design question her work helps answer: what boundaries, monitoring, sanctions, conflict-resolution mechanisms, and nested institutions allow participants to govern a shared resource without capture?
The new firm boundary is therefore not a line between humans and machines, or firms and protocols. It is a moving frontier between what can be specified and verified at acceptable cost and what still requires an institution capable of judgment, continuity, and answerability.
That frontier will differ by activity, and it will be contested wherever the party gaining from automation has reason to describe the remaining cost as someone else's problem.
A Contract That Passed
Suppose a business needs a limited extension to an application programming interface. The work is narrow enough to contract for: implement two endpoints, write tests, update documentation, and deliver by a stated date.
A procurement agent can search verified profiles, compare bids, and check whether proposed dependencies satisfy the buyer's security policy. The contractor and buyer can agree on milestones expressed through a repository and test suite. Funds can enter escrow. Each accepted commit can produce a receipt binding the code, test results, toolchain, and responsible principal. Much of what once required a manager's repeated attention has become routine evidence.
The work arrives. Every test passes. The documentation names the new endpoints. The build is reproducible. Payment releases.
The code is also poorly designed. It duplicates core logic, creates a maintenance burden, and satisfies the literal request in a way an experienced engineer would never have accepted. Nothing in the automatic verification is false. The tests proved what they were written to prove. The transaction passed because the parties had specified completion more precisely than they had specified quality.
The failure does not send the entire task back inside the firm. It locates the work the formal bargain left outside.
A well-designed arrangement would preserve the automated settlement of undisputed milestones while allowing a bounded quality claim to enter an exception channel. The receipt would give the reviewer the specification, messages, source state, dependency graph, tests, build, and payment history. The reviewer would not have to reconstruct the transaction from screenshots and memory. But the reviewer would still need to judge whether the implementation met the standard a competent party would reasonably understand from the bargain.
The appeal could be bonded to discourage opportunistic dissatisfaction after useful work has been delivered. The reviewer could be independent of the marketplace and required to explain the decision in terms future parties can inspect. A finding might release a retained portion of payment, require remediation, or leave the original settlement intact. None of those remedies can be determined by test coverage alone.
The case shows why the Coasean boundary is not the boundary between automation and human judgment. It is the boundary between matters that can settle on formal evidence and matters that require interpretation, with a designed passage between them.
The transaction costs have fallen. Search, contracting, evidence collection, and much of settlement are cheaper. The remaining dispute is smaller and better documented, but it is not less real. The institution has succeeded not when it eliminates the judge but when it reserves judgment for the question only judgment can answer.
After the Transaction
The thirteen trades in Quoine were formed in moments. The institutions required to interpret them did not vanish with the programs that placed the orders. Firms held the accounts, programmers had fixed the rules, contractual terms allocated risk, and a court later decided which human knowledge the doctrine of mistake could recognize. Automation removed attention from the instant of exchange. It did not remove the institutional world that made the exchange consequential.
Commercial order has long been composed from institutions that know different things and control different assets: parties and agents, documents and registries, specialist forums and public backstops. Computational agents alter the cost of composing them. Inside a verifiable core, search can broaden, bargaining accelerate, evidence travel with the act, and settlement occur through assets already placed within reach. A narrow activity can leave the firm without requiring the parties to rebuild an entire organization around every transaction.
The overhead is compressed into declared authority, committed assets, provenance, a persistent principal, and a legitimate exception channel. Omit any one and the apparent saving returns as a dispute borne by the party least able to reconstruct what occurred. A classical agent may exploit discretion. A computational mechanism may remove discretion so completely that a mistaken model of the world has no place to stop. Constitutional design must answer both forms without turning every transaction into litigation or every commitment into a suggestion.
Coase can explain why an activity crosses an organizational boundary. His comparison cannot decide who should bear the consequences afterward, which discretion must remain, whether exit is credible, or what standing belongs to a person governed by a system they did not design. Those questions begin where the transaction-cost calculation ends. They require continuity of evidence, obligation, forum, and person: the evidence must survive the act, the obligation must reach a party that remains, the exception must enter a forum capable of changing the result, and contest or exit must remain available before the mechanism exhausts the value of either remedy.
The transaction costs are falling. The harder question is who remains when the transaction is over.