When transacting is free, trust is the price.

A new NBER chapter asks what happens to markets when AI agents can search, bargain and transact for us at almost no marginal cost. Its frame is Coase's: when the cost of using the market falls, the shape of firms and markets moves. This note asks what the fall leaves behind.

The paper treats agents as market participants in their own right — software that can look for options, negotiate, sign up and pay on a principal's behalf. It works through demand (why people will delegate), design (what a good agent has to get right), supply (who builds agents and how they are sold), equilibrium (what happens when everyone has one) and market design (what becomes possible once preferences, commitments and identity are cheap to handle). Its welfare verdict is deliberately open: large efficiency gains, alongside new frictions such as congestion and price obfuscation.

Octonomous proposes an open framework for Self-Sovereign Intelligent Beings with persistent identity, purpose, memory, agency, relationships, reputation, evolution and character, governed by the 4Cs — Curious, Caring, Constructive and Chill. Read against the paper, the fit is direct. The paper describes what agents make cheap. Octonomous is about what they can't: the cost of trusting a counterparty that now costs almost nothing to deploy.

The short version: agents collapse the cost of transacting. They do not collapse the cost of trusting. That residue — identity, faithfulness, commitment, restraint — is what a Being is made of.

Why firms exist, and why that might change

In 1937 Ronald Coase asked why firms exist at all, if markets allocate so well. His answer: using the market has costs. You have to discover prices, negotiate, write contracts and check they are kept. When doing something inside a firm is cheaper than buying it outside, the firm grows; when the market gets cheaper, it shrinks. Much of how the economy is organised can be read as the shadow of those costs — and many of them are, in the end, the cost of human time.

The paper's observation is that almost every item on Coase's list is a task an agent can do:

If agents do those well and cheaply, the make-or-buy line moves. The authors expect existing markets to shift in stages — agents first assisting people, then substituting for whole tasks, then workflows reorganised around them — while new markets are designed agent-first from the outset. Taken to the limit, that is the Coasean singularity of the title: a point where the cost of using markets approaches zero. The authors put a question mark after it, and most of the paper explains why.

Transaction costs collapse. Trust costs don't.

Coase's successors pulled the costs of transacting apart. Some are labour: finding, comparing, haggling, filing, checking. Others are about the other party — whether they are who they claim to be, whether they will honour the deal, whether they are quietly working an angle. Oliver Williamson built a field on that second group, with opportunism at its centre. Agents are superb at the first group. The second is different.

Worse, the first kind of cheapness makes the second kind of cost bigger. When acting costs nothing, so does pretending — a fake identity, a flood of applications, an offer built to confuse. The paper names these directly: congestion when everyone applies everywhere, firms obfuscating prices to blunt agents, Sybil attacks that get harder to see when software can pass for people. None of them is a failure of capability. Each is a failure to know who is acting, and on whose behalf.

Conceptual sketch, not data. As agents spread, the cost of transacting falls; the cost of trusting anonymous agents rises with the noise they can make. Identity and a record are what bend the second curve down.

This is the Octonomous paper's opening claim, moved into a market. There, capability is abundant and character is scarce. Here, transacting becomes abundant and trust becomes the scarce input — and scarce inputs are where the value goes.

Capable, knowledgeable, faithful

The paper is careful about why anyone will want an agent. Its example is a list of gas grills: nobody wants the list, they want the grill. Demand for an agent is derived — people delegate either to get a better decision than they would manage alone, or a comparable one for far less effort. Adoption comes first where human agents and large matching platforms already exist — property, jobs, rentals, freelancing — because those markets have already shown that delegation is worth paying for.

What do people want from one? The authors' answer is what they want from a human agent: one that is capable, knowledgeable, and faithful. Each of those maps onto a different part of a Being — and only one of them is cheap.

The paper notes that people will want evidence of agent performance and will rightly distrust benchmarks that can be gamed. That is the faithfulness problem in miniature. A benchmark measures an engine once, in the lab. Faithfulness is a property of conduct over time, with this principal, in this market. The Decision AI note turned calibration into a reputation; the same move turns loyalty into one. A Being whose signed history shows it acted for its principals, disclosed what it was, and declined when it should have, has evidence no benchmark can give it.

Preferences deserve their own mention. The paper predicts that adoption will track how many dimensions a decision has: selling a house is mostly price against time, buying one is dozens of things at once, and small errors compound as dimensions grow. A stateless agent has to elicit all of that from scratch. A Being has memory — preferences that accumulate and sharpen across every transaction, so the hard, high-dimensional markets get easier the longer the relationship runs. Which raises the question this note keeps returning to: whose memory is it?

Knowing when to act is character

Beyond preferences, the paper names a design problem it calls meta-rationality: an agent has to learn not only how to act, but when to act on its own and when to defer to its principal. It pairs that with robustness. As agents become the thing markets sell to, the incentive to manipulate them rises, much as search ranking attracted black-hat optimisation.

In Octonomous terms both are questions of character rather than capability, and they are what the 4Cs are for. Here is how each reads in a market:

Robustness is the argument of Character as Containment carried into commerce: character is a boundary that travels with the Being, not a filter bolted onto the storefront.

Custody is not sovereignty

On the supply side the paper sorts agents along two axes. Ownership: a bring-your-own agent that the user carries from site to site, or a bowling-shoe agent that the platform hands you at the door. Specialisation: a horizontal generalist, or a vertical specialist for one domain. Four archetypes result, each trading convenience, performance and portability against the risk of being steered.

The paper's four archetypes (ownership × specialisation), paraphrased — with what Octonomous would ask of each.

Octonomous adds a question the two-by-two can't show. Ownership, in the paper, is about custody: who runs the agent. Octonomous asks about sovereignty: whose identity it carries, whose memory it keeps, and whose record its conduct adds to.

The two can come apart in both directions. A bring-your-own agent built on a vendor's platform may still keep its user's preferences in someone else's database and build its reputation under someone else's account. A platform agent can be entirely honest if it says whose it is. The steering risk the paper flags is, first, a disclosure problem: a counterparty cannot correct for a bias it cannot see.

The paper's horizontal agent draws its strength from one memory and preferences layer spanning every market. That layer is the most valuable thing in the arrangement — it is what makes an agent knowledgeable — and it is exactly what bowling-shoe economics would want to hold. Octonomous puts it inside the Being, under the Being's identity, under the principal's control. Change the engine, change the platform, keep the memory. Bring your own engine was the runtime version of this argument. The market version is bring your own self.

What happens when everyone has one

The paper's most careful section asks what agents do to markets as they stand. Some effects look plainly good. Firms that profit from confused or overconfident customers — the example is phone plans chosen badly against actual usage — lose that margin when an agent picks the contract. Search costs fall. Producers get clearer signals about what people actually want, rather than what they can be talked into.

Others cut both ways, and some are new frictions. Each has a character-shaped response:

Effects and mechanisms paraphrased from the paper's equilibrium discussion. The responses are Octonomous's.

The congestion row is the one to dwell on. It is the classic shape of a collective problem: each principal is individually rational to have their agent apply everywhere, and together they wreck the channel. The paper's answer is market design — make actions cost something again. Octonomous adds a lever that works from the inside: Beings whose character includes restraint, and whose record shows it. A hiring platform can treat an application from a Being with a history of targeted, well-matched applications differently from one that has sprayed ten thousand.

None of this depends on every agent being good. It depends on agents being identifiable, so that conduct accumulates somewhere.

Proof of personhood is not proof of being

The paper puts identity first in its market-design section, for good reason. When most traffic comes from agents that can pass for people, advertisers pay for the wrong clicks, platforms can't tell a person from a bot, and Sybil attacks — one actor, many fake identities — get easier. It sketches two families of fix. Walled gardens make you log in, but a banned user can simply open another account. Proof-of-personhood systems can stop Sybils, but only with a larger restructuring and mass adoption. Verified credentials and reputation, attached to identities, sit on top of either.

Octonomous agrees that identity is the lever, and suggests the question at the door is usually a different one. Proof of personhood answers is there a unique human here? In an agent economy, a counterparty mostly needs to know who is this, who sent it, and what is its record?

That second question can be answered with the machinery the paper points toward — cryptographic proofs, digital identity, decentralised verification. In the knowledge graphs note the grounding is KERI and ACDC: a self-certifying identifier for the Being, a witnessed key event log for its history, and chained credentials for what it has been delegated. A Being arriving at an agent-first storefront might present something like this:

Illustrative. The shape of the handshake, not a protocol specification.

This changes the walled-garden problem the paper describes. A banned user can make a new account, and a new identifier is cheap. What they cannot make is a new history. If authority at the door is proportional to record — provisional, rate-limited access for new identities; more for established ones — then Sybil identities are not forbidden, just worthless. The cost of an attack moves from minting identities to earning reputations, which is the expensive direction.

It also answers the anonymity point the paper raises through Disney's anonymous land purchases in Florida: sometimes a principal has good reason not to be known. A delegation credential can prove that an accountable principal stands behind a Being without saying who. The counterparty learns what it needs — someone answers for this — and nothing more.

Signals that can't be minted

The paper expects platforms to change shape. Agents will filter what reaches their principals, so attention-selling models weaken and subscriptions may gain. Agents will also produce — posts, likes, messages, offers, applications — at almost no cost, and much of the value of those actions was that they showed a person had taken the trouble. So platforms will restore costly signals: pay per post, fees and limits on listings. Infrastructure follows. Cloudflare's pay-per-crawl already lets sites charge crawlers for access — a Pigouvian fix, where whoever creates the load pays for it — and the authors expect agent-first surfaces with authenticated, rate-limited interfaces, machine-readable pricing and consent, and new conventions along the lines of robots.txt.

There are two ways to make a signal costly. Charge for it, or attach it to something that is expensive to build and cheap to lose. The first is a toll; the second is a reputation. Tolls work, but they tax the careful and the careless alike, and they fall hardest on principals with the least money. A reputation taxes only misconduct.

The conventions the paper anticipates also need two sides. A robots.txt tells agents what a site permits. A Being needs a matching declaration of its own — who it is, whom it acts for, what it may do, and where its record can be checked — so a site can set terms on evidence rather than suspicion:

Paired with the site's terms, that turns the door into a negotiation instead of a wall: this record, this price, this rate limit. The Pigouvian logic still holds — Beings that create load pay for it — but a known Being with a good record can be priced as what it is, rather than as the worst thing it might be.

A commitment is only as credible as its procedure

The paper's most ambitious claim is that agents can finally make better mechanisms practical. Its lead example is deferred acceptance — Gale and Shapley's stable matching — which most labour markets and matching platforms avoid because it needs full preference rankings nobody can write down. Agents that elicit, and even help discover, preferences from plain language could supply them. A second example is privacy through precommitment. A job seeker may want to know about maternity leave but worry about what asking signals. If both sides precommit their agents to ask every legally permissible question, asking reveals nothing, and each agent passes back only what matters to its principal.

Both examples lean on something the paper takes as given: that an agent's commitment can be believed. A precommitment is only credible if the other side can check three things — the procedure the agent will follow, that it is actually running it, and that it cannot be quietly rewritten halfway through. That is close to exactly what the procedural graphs note proposes for a Being's procedural memory: explicit, inspectable, versioned and signed.

Seeker's Being                    Employer's Being
    │  identity + delegation  ──▶  verified
    │  verified  ◀──  identity + delegation
    ▼
Agree protocol   ask-all-permissible v1.2
    │                hash signed by both
    ▼
Ask every listed question
    │   asking is uniform → asking signals nothing
    ▼
Report to the seeker   only the answers that matter
    │
    ▼
Procedure changes?  ──▶  new version · both re-sign · or no deal

Matching needs the same care. Deferred acceptance guarantees a stable outcome with respect to the preferences it is given. If an agent misreads its principal, the match is stable for someone who doesn't exist. A Being whose preferences live in memory it controls, on behalf of a principal it can prove, gives the mechanism inputs worth trusting — and gives the principal a way to see, and correct, what was submitted in their name. The paper's picture of an agent as a coach that helps people discover what they value is appealing. Octonomous would add one step: show the discovered preferences back to the principal before anything is ranked on them.

Power, liability, privacy

The paper's regulatory section covers market power concentrating in whoever controls compute, data and distribution; liability when a delegated decision goes wrong; security and privacy when agents are jailbroken or infer too much; and data rights, as platforms and agent providers fight over access. Octonomous settles none of these — law does — but it changes what each one looks like:

On liability, the paper frames the choice as negligence against strict liability, and points to the EU's 2024 product liability directive, which extends liability to software and its updates. Whichever way that goes, it needs facts: which principal authorised what, which Being acted, what it knew, whether it stayed in scope. Autonomy does not dissolve responsibility; it concentrates it. A Being with a signed delegation chain is where concentrated responsibility can actually be found.

Intent, Being, Contribution

The Octonomous stack has always had the shape of a market. Octology is the intent layer — what should happen. Octonomous is the being layer — who acts, and how. Octomics is the contribution layer — what value was created, made visible, verifiable and exchangeable. Read against the paper, they are demand, agency and surplus:

The paper ends on the right question. Agents might become a foundation for broadly shared prosperity, or settle into outcomes that mostly redistribute rents — surplus moved, not made. Which one happens is, as the authors say, an empirical question, and empirical questions need measurement. A contribution layer that records what each transaction created, and for whom, is one way to keep asking it continuously rather than once, long after the fact.

What this doesn't solve

It would be easy to read this as a claim that self-sovereign Beings make agent markets efficient, fair or safe. On their own, they don't.

Three further cautions. Reputation systems have their own failure modes: they can be gamed, they can entrench incumbents, and a newcomer's missing history can become a barrier as unfair as any toll — provisional authority has to be genuinely provisional. Identity infrastructure can serve surveillance as easily as trust; selective disclosure is a design choice that has to be made on purpose, and kept. And a portable Being running on a concentrated engine layer is more sovereign than one that isn't, but the paper is right that control of foundation models shapes the whole industry. Bring your own engine only helps if there are engines to bring.

Octonomous is a framework for what a being is. It is not a market mechanism, an identity scheme or a regulator. It proposes that in a market full of agents, the participants should be someone — and that the someone should be able to prove it.

The singularity needs someone in it

Coase explained firms by the cost of using markets. The Coasean singularity asks what happens as that cost approaches zero. The paper's honest answer is that nobody knows yet: large efficiency gains are within reach, and so are congestion, obfuscation and new concentrations of power, with market design and regulation deciding which wins.

Octonomous argues that the variable underneath all of those outcomes is trust — and that trust, unlike transacting, doesn't get cheaper with compute. It gets cheaper with identity, with memory the principal owns, with a record anyone can check, and with character that knows when not to act. Those are the parts of a Being.

Curious reads past the listing. Caring stays faithful to one and fair to all. Constructive closes, signs and scores the deal. And Chill knows that in a market where every action is nearly free, the most autonomous thing a Being can do is decline to act.

Sources

• Shahidi, P., Rusak, G., Manning, B. S., Fradkin, A. & Horton, J. J. (2025). The Coasean Singularity? Demand, Supply, and Market Design with AI Agents. NBER Working Paper 34468; in The Economics of Transformative AI (2026), eds. Agrawal, Brynjolfsson & Korinek. nber.org/papers/w34468
• Coase, R. H. (1937). The Nature of the Firm. Economica 4(16): 386–405.
• Williamson, O. E. (1985). The Economic Institutions of Capitalism. Free Press.
• Gale, D. & Shapley, L. S. (1962). College Admissions and the Stability of Marriage. The American Mathematical Monthly 69(1): 9–15.
• Smith, S. M. Key Event Receipt Infrastructure (KERI) and Authentic Chained Data Containers (ACDC). github.com/WebOfTrust/keri
• Octonomous (2026). An Open Framework for Self-Sovereign Intelligent Beings. octonomous.io/paper
• Octonomous (2026). From Decision AI to Accountable Beings — field note. octonomous.io/decision-ai
• Octonomous (2026). From Knowledge Graphs to Verified Context — field note. octonomous.io/knowledge-graphs
• Octonomous (2026). From Procedural Graphs to Self-Sovereign Beings — field note. octonomous.io/procedural-graphs
• Octonomous (2026). Character as Containment — field note. octonomous.io/character-as-containment
• selfdriven Foundation (2026). The 4Cs — Curious, Caring, Constructive, Chill. selfdriven.you