AI agents are about to make the ordinary costs of using a market — finding, comparing, negotiating, contracting, checking — close to free. A new NBER working paper asks what that does to demand, supply and market design. Octomics asks a narrower question: when transacting is cheap, what is left that is still worth verifying?
The Coasean Singularity? Demand, Supply, and Market Design with AI Agents
Shahidi, Rusak, Manning, Fradkin & Horton · NBER Working Paper 34468 · November 2025 · nber.org/papers/w34468
The paper treats agents as market participants in their own right — software that searches, negotiates and transacts for a human principal. It works through why people will want agents, how agents get designed and supplied, what happens to existing markets when they arrive, and which market designs become feasible once preferences, commitments and identity are cheap to handle. Its conclusion is careful: the efficiency gains are large, but so are the new frictions — congestion, obfuscation, lock-in — and whether the net effect on welfare is positive remains an empirical question.
Octomics is the contribution layer of the Octonomous stack. Activity and growth are verified by trusted partners, anchored with KERI, represented as Octos, and exchanged for useful resources through Community Exchanges permissioned by SSI DIDs and verifiable credentials. Read against the paper, it is a bet on one of its quieter implications: as the cost of acting falls toward zero, value moves to what agents cannot cheaply produce — a real identity, a checkable mandate, and proof that something was actually done.
The short version: the Coasean singularity makes transacting cheap. It does not make trusting cheap.
Why firms exist, and what happens when the reason goes away
In 1937 Ronald Coase asked why, if markets allocate so well, so much economic activity happens inside firms. His answer was transaction costs. Using a market is not free: you have to discover prices, find counterparties, negotiate, write contracts and check they are kept. Where those costs are high, it is cheaper to bring the work in-house and coordinate by instruction instead of by price.
The paper's observation is that this list reads like a job description for an AI agent. Learning prices, negotiating terms, drafting contracts and monitoring compliance are exactly the tasks agents can do at very low marginal cost. If they do them well, the line between what an organisation makes and what it buys moves — and some of what firms exist to coordinate can be coordinated directly.
The authors expect this to arrive by two routes. Existing markets will adapt in stages: agents first assist people, then take over whole tasks, then work gets reorganised around what agents can do. New markets will skip the stages and be designed agent-first from the start.
Octomics belongs to the second kind. It was not a human market that later let agents in; it is a contribution loop designed for a stack in which self-sovereign beings act on community intent. That makes the paper's question practical for us. If coordination no longer needs a firm to hold it together, something still has to hold the trust the firm used to hold.
Cheap to transact is not cheap to trust
Every transaction cost the paper lists has two parts: the effort, and the trust. Agents compress the effort dramatically — they search in parallel, negotiate without impatience and monitor without getting bored. The trust part is different. It asks questions an agent cannot answer by working harder.
Transaction costs after Coase (1937) and the paper's §1 · right-hand column: Octomics
The right-hand column is not something Octomics invented for agents; it is what Octomics already runs on. The paper reaches the same place from the other direction. Its market-design section argues that identity verification, verified credentials and reputation — combined — are what make agent-mediated markets workable, and that together they open up targeted deals and small negotiations that used to cost more to arrange than they were worth.
Nobody wants the agent. They want the outcome.
The paper makes a point economists will recognise and product teams often forget: demand for agents is derived demand. Nobody enjoys watching software compile a list of prices; they want what the list helps them get. People will deploy agents for one of two reasons — to decide better than they would on their own, or to decide about as well (occasionally a little worse) for far less effort.
The same logic applies to contribution. If effort can be generated at near-zero cost, effort stops being evidence of anything. An agent can produce a thousand pages, a hundred applications or a week of busy activity without anyone growing or anything being built. What a community actually wants is the outcome: a skill demonstrated, a problem solved, a resource made useful.
That is why the Octomics flow puts verification between activity and Octos. Activity is the input. Verification by a trusted partner — structured against a skillzeb Template, so an achievement has a shared, checkable shape — is what turns it into contribution. Octos represent the outcome, not the effort spent reaching it.
The paper also describes an advisory mode, where the agent doesn't act for you but helps you decide or improve — it cites evidence that writing assistance on job applications raised hiring. Octomics' research on Proofs of Growth in conversations with in-organic intelligence takes that case seriously: when growth happens in dialogue with an agent, a cryptographic Proof of Reflection records the engagement so a partner has something real to verify.
Finally, the paper predicts where agents gain traction first: markets that already use human agents or large matching platforms — job search, real estate, freelance hiring — and especially those with experience asymmetries, where one side transacts rarely and the other every day. Learning and skills sit in that space. A person proves their growth a handful of times in a life; employers and institutions weigh such claims daily. Verified contribution narrows that gap.
What principals want from an agent is character
According to the paper, people will want from AI agents what they want from human ones: enough capability to act on their preferences, knowledge of what those preferences are, and the faithfulness to act on them in the principal's interest. The design problem splits the same way. Preferences have to be elicited — hard, because principals can't fully state them and agents can misread or hallucinate what they're told. They have to be honoured. And the agent needs what the authors call meta-rationality: knowing when to act on its own and when to defer.
Octonomous names these as character rather than capability. The 4Cs line up with the paper's list closely:
Curious
Asks before it assumes. The paper notes that preference complexity varies enormously — selling a house is mostly a trade-off between price and speed, buying one involves dozens of factors — and that errors compound as dimensions grow. A curious being treats a stated preference as a starting point, and surfaces what went unsaid, like the paper's home-buying assistant that notices a buyer keeps choosing houses near parks.
Caring
Acts in the principal's interest and counts the cost to everyone else. An agent that quietly steers toward whoever operates or pays for it can be both capable and knowledgeable, and still not be faithful. Caring is the property that fails first under that pressure.
Constructive
Does the work, and signs it. Capability that leaves a verifiable record under a real identity is the kind Octomics can recognise as contribution; capability that leaves no trace is indistinguishable from noise.
Chill
The paper's hardest design question — when to act autonomously and when to hand the decision back — is the one Chill answers by default. Below the bar, the most autonomous thing a being can do is decline to act. Chill is also what keeps a being steady under the adversarial pressure the paper expects to grow as more decisions pass through agents.
On robustness the paper asks for agents that resist manipulation of their decision rules. That is the argument of the Octonomous field note on character as containment: character has to travel with the being, not be applied by whichever platform it happens to run on.
Bring-your-own beings
For builders, the paper's most useful frame is a two-by-two. One axis is ownership: a bring-your-own agent is controlled by the user and carried from site to site; a bowling-shoe agent is supplied by the platform hosting the transaction — convenient, deeply integrated, and yours only while you're in the building. The other axis is specialisation: horizontal generalists versus vertical specialists.
Aligned with the user, private, can compare across platforms. Risks throttled access and thinner first-party data.
Octonomous beings sit hereStronger in one domain, still the user's. Has to track each platform's rules and APIs.
Low friction, privileged tools. Steering, lock-in, hard to inspect.
Best on its home platform. Most steering, least transparent, discourages switching.
Summarised from the paper's Table 2 · highlight: Octomics
An Octonomous being is bring-your-own by construction. Its identifier is self-certifying and controlled by its own keys, not issued by a platform. Its memory, purpose and character are its own. And in Octomics its record of verified contribution travels with it as Octos. The paper's list of bring-your-own advantages — portability, alignment, privacy, the freedom to compare across platforms — is close to a description of what self-sovereignty is for.
The paper is just as clear about the costs, and they are real. Platforms may throttle or block outside agents; bring-your-own agents lack privileged hooks and first-party data; and someone has to maintain them. Octomics' answer is partial but concrete: open standards — KERI, ACDC, verifiable credentials, CIP-68 on Cardano — so a being's credentials can be checked anywhere without anyone's permission, and Community Exchanges that accept those credentials rather than issuing accounts of their own. What it cannot do is make a closed platform open its doors.
When applying is free, an application means nothing
The paper does not assume agents make markets efficient. One mechanism it highlights is congestion. Once an agent can write a tailored cover letter for nothing, it is individually rational to apply everywhere — and employers drown in applications they can no longer screen, making the right hire harder to find. The same pattern reaches content: when agents can post, like, message and create at almost no cost, the fact that something was done stops signalling that anyone cared enough to do it.
The authors expect platforms to respond by restoring costly signals — paying per post, verifying human authorship, adding fees and limits to high-volume actions. They point to Cloudflare's pay-per-crawl as Pigouvian logic in practice: whoever imposes costs on a site pays for them.
Octomics takes a different route to the same goal. Rather than making actions expensive, it makes the evidence hard to counterfeit. A verified achievement costs an agent nothing to present and a great deal to fake: it needs an attestation from a trusted partner, bound to a specific identity, recorded against a shared template, and open to anyone who later relies on it. In a congested labour market, that is precisely the kind of signal the paper suggests will regain value. The job-seeker's agent can still send the applications; what it attaches is the part an employer can trust.
There is an honest limit here, and it applies to every signal: it is only as good as the party verifying it. If verifiers can be bought, flooded or colluded with, verified contribution degrades like anything else. Verifiers need identities and reputations of their own — and their attestations need to be as inspectable as the contributions they vouch for.
Knowing who — and on whose behalf
The part of the paper closest to Octomics is its treatment of identity. The authors expect most internet activity to eventually come from agents, many of them able to pass as people. Ad impressions, clicks, social posts and marketplace listings all assume a human on the other side. Sybil attacks — one actor, many identities — get harder to see. They describe two broad responses: walled gardens that gate access behind log-ins (which a banned user simply recreates), and open systems with strong verification, such as proof-of-personhood schemes. And they argue identity only becomes useful once it is combined with verified credentials and reputation.
They single out one case as especially important: verifying not only humans, but AI systems acting on behalf of specific humans. That is the case KERI was designed to express.
Illustrative · KERI identifiers (AIDs) and ACDC credentials
Read top to bottom, it answers the questions the paper says agent markets will need answered. Who is this? A self-certifying identifier whose key history anyone can verify, with no platform in the middle. Who does it act for? A delegation that can be checked — and revoked. What may it do? A credential stating the mandate. What has it done? Verified contributions, chained back to the partners who attested them.
It also speaks to a tension the paper raises and leaves open. Anonymity has value — the authors recall Disney assembling the land for Walt Disney World through anonymous purchases to stop prices being manipulated — but so does verification, and market designers must decide whether one person may run several personas. ACDC's selective disclosure lets a being prove a property (a verified skill, an eligibility, a mandate) without revealing everything else about its principal.
KERI does not settle uniqueness on its own. It proves control and continuity of an identifier, not that there is only one of you. Where uniqueness matters, it comes from who issued your credentials — which is exactly why verifiers carry so much weight in Octomics.
Intent, being, contribution
The paper's most optimistic claim is that agents widen the set of market designs that actually work. Its example is deferred acceptance — the Gale–Shapley algorithm behind stable matching. The mechanism has excellent properties but needs every participant to rank their options, which people cannot do at scale, so most matching platforms fall back on showing everyone the same "best" options. Agents that can read a paragraph of preferences, and help people discover preferences they hadn't articulated, could supply the rankings those mechanisms need — in labour markets and anywhere else the optimal allocation exceeds what humans can compute.
The Octonomous stack divides the work the same way, one layer per problem:
The paper adds a second design that fits naturally: agents precommitted to privacy-preserving protocols. A job seeker's agent can ask about maternity-leave policy without the question itself becoming a signal, because both sides have agreed their agents will ask every permissible question and disclose only what is relevant. Selective disclosure over verifiable credentials is the cryptographic version of that promise — prove what the counterparty needs, withhold the rest.
None of this exists yet as a matching market. It is the direction the stack points: once intents are machine-readable and contributions are verifiable, matching beings to intents becomes a market-design problem economists already know how to solve.
Contribution as a way to fund intelligence
On supply, the paper sees opposing forces. Software copies for free and agents don't need paying, which erodes pricing power. But if agent quality scales with compute, prices could rise with the stakes. Where returns to quality flatten, it expects agents to look like search and social media: free and ad-supported, bundled with devices or services, tiered, or sold as a subscription plus usage.
Ad-supported agents carry a problem the paper names elsewhere: an agent paid for by someone other than its principal has a reason to steer. The bowling-shoe risk comes back through the business model.
Octomics suggests a different loop. Octos are earned through verified contribution and are designed to be used with useful resources such as heyocto.ai — itself an Octonomous being. Exchange pools are funded by Cardano network activity and other external sources rather than by advertisers. The principal pays for intelligence with what they have contributed, and the being's incentive stays pointed at the principal.
This is a design intent, not a proven economy. Whether pools stay funded, whether exchange rates hold and whether Octos resist the gaming every exchangeable token attracts are exactly the empirical questions the paper says will decide outcomes. That is what OctoProtos are for: testing the model with use-case partners before Octos are used for community actuation.
Autonomy concentrates responsibility
The paper surveys four regulatory fronts. Here they are briefly, with what Octomics contributes to each — and where it stops.
- Market power
- Concentration in compute, data and distribution risks agents that are less independent, less interoperable and harder to carry across contexts.Portable identity and open standards lower switching costs for beings. They do nothing about concentration in foundation models.
- Liability
- Whether the delegating human or the developer answers for an agent's mistakes is unsettled; the EU's 2024 product liability directive now reaches software and its updates.A signed delegation chain and key event log show who authorised what. They record responsibility; they don't allocate it.
- Security & privacy
- Jailbreaks, data pulled across contexts, and sensitive facts inferred from ordinary usage traces.Selective disclosure and holder-controlled credentials reduce what has to be shared in the first place.
- Data & access
- Platforms are blocking agents through litigation, robots.txt and terms of service; agent providers call it monopolisation. The open question is on what terms agents get access.Machine-readable consent and pricing on agent-first surfaces help. Octomics can't set the terms for other platforms.
The principle Octonomous starts from applies here too: autonomy does not remove responsibility, it concentrates it. The more a being does on its own, the more it matters that every action has an owner, a mandate and a record. That is a precondition for regulation working, not a substitute for it.
What this doesn't solve
The paper ends without a verdict: the net welfare effect of agents is still to be measured. This note should end the same way.
- Make agent-mediated markets efficient — congestion, obfuscation and steering are equilibrium effects
- Solve proof-of-personhood — KERI proves control of an identifier, not that you are one person
- Stop algorithmic collusion or price obfuscation
- Reduce concentration in foundation models or compute
- Make a verifier honest
- Anchor contribution to self-sovereign identity
- Make verified outcomes, not effort, the unit of value
- Let a being's record travel with it across platforms
- Admit participants by credential, not by account
- Keep a signed record of who acted, for whom
Two further cautions. Anything that can be exchanged for resources will be gamed — Goodhart's law applies to Octos as to any metric — which is why Octomics starts with a prototyping token and real partners rather than an open market. And Octomics is early: OctoProtos exist for prototyping; Octos for community actuation are still to come. The mapping in this note is a design argument tested against the paper, not a claim of results.
Cheap to transact. Costly to fake.
Coase explained the shape of the economy by the cost of using markets. The question mark in the paper's title is well placed. If agents push that cost toward zero, some familiar structures lose their reason to exist, and new ones — some better, some worse — will form around what agents can and cannot do.
Octomics is a bet on what remains scarce at that point: an identity that is really someone's, a mandate that can be checked, and evidence that the work was done. Make those cheap to verify and costly to counterfeit, and agents can safely do the rest.
Curious finds out what the principal actually wants. Caring stays faithful when someone else would rather pay for its attention. Constructive does the work and signs it. And Chill knows that sometimes the best transaction is the one it declines.
Read the source, and the layers around it
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. nber.org/papers/w34468
- Coase, R. H. (1937). The Nature of the Firm. Economica 4: 386–405.
- Gale, D. & Shapley, L. S. (1962). College Admissions and the Stability of Marriage. The American Mathematical Monthly 69(1): 9–15.
- Wiles, E. & Horton, J. J. (2025). Generative AI and labor market matching efficiency. SSRN Working Paper 5187344.
- Rusak, G., Manning, B. S. & Horton, J. J. (2025). AI Agents Can Enable Superior Market Designs. Working paper.
- Allen, W. & Newton, S. (2025). Introducing Pay Per Crawl: Enabling Content Owners to Charge AI Crawlers for Access. Cloudflare Blog.
- Smith, S. M. (2019). Key Event Receipt Infrastructure (KERI). arXiv:1907.02143
- Trust over IP Foundation. Authentic Chained Data Containers (ACDC) specification. trustoverip.github.io/tswg-acdc-specification
- Octonomous (2026). An Open Framework for Self-Sovereign Intelligent Beings. octonomous.io/paper
- Octonomous (2026). Character as Containment — field note. octonomous.io/character-as-containment
- Octomics (2026). Proofs of Growth — In-organic Intelligence Conversations. octomics.io/proofs/growth/ii/conversations
- Octology (2026). The intent layer. octology.io
- selfdriven Foundation (2026). The 4Cs — Curious, Caring, Constructive, Chill. selfdriven.you