When markets start to be negotiated by machines: the Agent-Mediated Economy and the next economic order

TL;DR

B2A asks whom a business sells to. The bigger question is what the market itself becomes when buyers and sellers both take part through agents. Technology can lower coordination costs; it cannot decide on its own who captures the gain.

• Agents are a new drop in transaction costs. The internet cut search costs, but evaluation, negotiation, execution and accountability still rest mostly on people. Agents can step into all of these at once. The real change is not that every trade becomes automatic; it is that many trades that never happened because coordination cost too much become economically viable.
• But transaction costs do not fall to zero. Agents add new ones: model calls, execution errors, identity verification, fraud prevention and assigning liability. A new way of organizing only lasts if what it saves keeps exceeding what it adds.
• The most important institutional question is whom the agent represents. Once paid promotion is built into task planning and tool selection, conflicts of interest are harder to spot than in a traditional ad. Whoever defines the agent objective decides whether it is the consumer agent or the platform distribution tool.
• When search stops being scarce, credibility becomes the scarce resource. Commercial information optimized for machines can be false too, and the falsehood can be automated. The key infrastructure is not better recommendation but verifiable economic facts, authorization and liability.
• The agent era is not necessarily a win for decentralization. More users, more merchants, more tools and higher completion rates feed a network effect that could make a few agent platforms the new centers of economic coordination. Efficiency and concentration can rise together.
• It is still early. Public agent traffic data keep showing that the overwhelming majority of requests stay on browsing and search pages, and only two or three percent reach checkout. Agents are already changing discovery; autonomous transactions lag well behind.

Check notes: the 2026 products and figures were checked one by one, and three were rewritten to match the original sources. (1) OpenAI launched Businesses in ChatGPT on November 10 as a beta, and the first Business Agents are open only to US e-commerce and online-service companies. (2) The Universal Commerce Protocol was launched by Google together with Shopify and other retailers, not pushed by Shopify alone. (3) The draft cited HUMAN Security for "agent browsing traffic up about eightfold from August to October, 85.2% on browsing and search, 1.9% reaching checkout"; the original report could not be found. The verifiable HUMAN figures are 77% on product and search pages and 2.3% on checkout pages for 2025 as a whole, and 3.16% on checkout and payment in the April 2026 monthly report. The scale matches the draft and its argument, so the body uses the verifiable numbers. Note also that HUMAN reports a different "eight times": automated traffic growing eight times faster than human traffic. The two are easy to confuse.

Introduction: future transactions may no longer need us to take part in person

Picture an ordinary morning a few years from now. You are moving to Japan for six months and need to rent an apartment, buy insurance, set up internet and put together a remote-work setup. You do not open a property site, compare insurance quotes one by one, or register on five or six platforms. You just tell your AI agent:

I will live in Tokyo for the next six months. Budget under 250,000 yen a month, no more than ten minutes on foot from a subway station, with stable internet and a setup that works for remote work. Arrange it, but every long-term contract needs my approval.

Your agent pulls listings from several property providers' agents, verifies the lease terms, compares insurance plans, checks network coverage and asks for executable quotes. The merchants' agents decide, based on inventory, fulfilment cost, customer conditions and trading rules, whether to offer the service and at what price. What the two sides exchange is no longer web pages but commercial terms, authorization credentials and commitments that a program can verify. What you receive at the end is not dozens of pages to read but a plan that has been checked, compared and readied for booking. You only approve the parts that carry long-term obligations.

This is not yet common, but it raises a question worth taking seriously: as people hand searching, comparing, negotiating and executing to AI agents, will markets keep working the way they used to?

In November 2026, Olivia Moore and Justine Moore of a16z proposed B2A (Business to Agent), arguing that AI agents are becoming a new kind of customer and a new distribution target for businesses. That captures a real shift, but it still uses the classic business-model question: whom does a company sell to? The question worth asking is a different one. If buyers and sellers both take part through agents, how do prices form, how does information travel, how are deals struck, how is trust built, and who holds market power?

That goes beyond B2A. I prefer the term Agent-Mediated Economy (AME) for the structure that may emerge. It is not a new economy separate from the existing one but a market coordination mechanism in formation. The core change is not that AI becomes a new consumer; it is that people delegate part of their economic decision rights and execution rights to software agents that run continuously. If that is widely adopted, we are seeing not just another migration of distribution channels but a reorganization of how markets are intermediated.

1. Markets were never only about buying and selling

To understand AME, start with an economics question: why do people need market intermediaries at all?

In the simplest model, buyers have demand, sellers have supply and prices coordinate them. Reality is messier. To buy a used car you have to find a seller, check the car, compare prices, negotiate the contract, arrange payment and live with whatever goes wrong after delivery. None of that produces a car, yet it uses up time, knowledge and resources.

In 1937, Ronald Coase argued in The Nature of the Firm that using the price mechanism itself has a cost, an idea that grew into transaction cost economics. Markets need firms, brokers, platforms, banks and all kinds of intermediaries largely because they reduce friction in trade.

Seen this way, centuries of commercial technology are a history of falling transaction costs: industrialization cut the cost of production and transport, telecoms the cost of distant communication, the internet the cost of spreading and finding information, e-commerce the cost of trading and paying online. Yet the internet never really solved one problem: information became easy to get, but understanding it, choosing and executing still relied mostly on people. The internet solved how information reaches people, not how it turns into action.

That is the institutional change agents may bring. The traditional internet provides an entry point for information; an agent may become the execution intermediary between expressing a need and taking economic action.

2. From an information economy to an action economy: another drop in transaction costs

Break the total cost of a trade down roughly:

total transaction cost = search + evaluation + negotiation + execution + monitoring and accountability

This is not a formula with a common unit of measure but a frame for analysing friction. Over the past twenty-odd years the internet pushed down the first term sharply; the others did not fall in proportion. You can find a hundred firms offering similar services in ten minutes and still spend hours working out which one is trustworthy and which actually fits; more complex trades add communication, bargaining, payment and repeated confirmation.

The potential value of an agent is that it can step into all of these at once: search suppliers, compare services against rules set in advance, call interfaces for quotes, check contract terms, and carry out some of the execution under explicit authorization. Interaction centred on moving information can extend into interaction centred on getting a task done. What a user buys may no longer be the right to use some software but a verifiable outcome. Someone unwilling to pay five travel-planning subscriptions might pay a service fee for one complete, reliable cross-border trip. Businesses may stop charging per seat and start charging per completed order, per verified task or per unit of work actually saved.

There is one important constraint: agents do not drive transaction costs to zero. They lower the cost of human search and execution but add model-call costs, execution-error costs, identity-verification costs, fraud-prevention costs and liability costs. A new way of organizing is only competitive in the long run if what it saves keeps exceeding what it adds. That is why agents will spread first in software development, standardized procurement and low-risk information work, while medicine, finance and complex legal transactions need more careful institutional design.

The real change is not that every trade is completed automatically by AI. It is that many trades that would never have happened, because coordinating them cost too much, become economically viable.

3. Hayek's knowledge problem: will agents make markets more efficient?

In 1945, Hayek argued in The Use of Knowledge in Society that one of the central problems of an economy is that knowledge is dispersed among countless individuals, and no central planner can hold everyone's needs, resources, preferences and local circumstances. The value of the price mechanism is that it uses dispersed information to coordinate activity without requiring anyone to understand the whole system.

That problem now has new technical conditions. An agent with ongoing authorization and long-term context could, in principle, understand some of a user's preferences better than a website does: budget constraints, past choices, delivery requirements, schedule. A supplier's agent connects to inventory, capacity, pricing, fulfilment rules and live operating data. When two agents coordinate, the information they exchange may be more structured than a web page and closer to the actual constraints of the trade. That could improve matching.

It does not mean markets will approach perfect information. Human preferences are not a fixed set of parameters. People often do not know what they want. Buying a car involves transport efficiency, but also identity, taste and an imagined future life; choosing a university involves fees and job prospects, but also an atmosphere that is hard to quantify. These preferences are situational, social and changeable, and even a rich history may not reveal them.

More importantly, preferences can be shaped by the system itself. When an agent filters your information, recommends your purchases and plans your life over the long run, it does not only observe your preferences; it changes how you form them. So AME is not only about allocative efficiency but about how preferences are formed. Markets used to influence consumers through advertising and brands; in future they may change consumer choices indirectly, by influencing consumers' agents. An agent is not a fully transparent, neutral tool for processing information. It is becoming a new interpreter of the economy and a new designer of choice architecture.

4. An easily overlooked question: whom does the agent represent?

This is the most important institutional question in AME.

The principal-agent problem is not new: shareholders delegate to managers, investors to fund managers, consumers to brokers. Whenever the principal and agent differ in goals, information or incentives, agency problems appear. AI agents widen that relationship further.

Suppose you ask a shopping agent to recommend a laptop and it gives you three brands. How do you know the recommendation serves you? Is it because performance, price and after-sales service suit you best, or because a merchant paid for placement, or because the agent platform has a deal with a brand, or simply because one merchant's data was easier for the model to retrieve and call? A traditional ad can usually be recognized as an ad. When promotion is built into an agent's task planning and tool selection, the conflict of interest is much harder to see.

Ideally, a consumer agent would optimize mainly for the user, within the bounds of lawful authorization. In commercial reality, the agent's provider is pursuing user satisfaction, platform revenue, advertising income and ecosystem expansion at once, and those goals do not always align. That raises a new question: when agents become the main intermediaries of trade, who has the right to define their objective function?

If users have real control, can move their data and can audit the agent's decisions, the agent is more likely to act for the consumer. If the platform controls the model, the user context, the point of entry and the commercial ranking rules, the agent may become a new distribution tool for the platform. Technically the two structures can look almost identical; economically they are fundamentally different. The first strengthens consumer sovereignty; the second may give more control over information and transactions than any search engine ever had. The more capable agents become, the more authorization boundaries, disclosure of commercial interests and explainable decisions matter.

5. From two-sided platforms to multi-agent markets: how the next platforms form

One of the most successful business forms of the internet era is the two-sided platform: marketplaces connect buyers and sellers, ride-hailing connects drivers and riders, payment networks connect merchants and payers. Rochet and Tirole's theory of two-sided markets analysed how platforms coordinate different participants through price structure and grow through cross-side network effects.

Agents will not automatically wipe out these platforms, but they may change how platforms organize supply and demand. A traditional platform builds its own interface, recommendation, reviews, search and transactions, and consumers have to enter the platform to use them. In future a user may first enter their own agent, which then calls the goods, services and transaction interfaces of different platforms. Platforms used to own the user entry point, and with it the power to allocate demand. In future the user agent may be the entry point, and product platforms may be reduced to callable networks of supply and fulfilment. Platforms would then compete not only for human attention but for agent calls.

Do not declare traditional platforms dead too soon. Those with exclusive inventory, real fulfilment capacity, credit systems and transaction guarantees keep strong advantages. An agent can lower the cost of finding a product, but it cannot conjure up warehouses, logistics, credit or supply. The platforms most at risk are those whose moat is mainly an information entry point and user switching costs, without distinctive supply.

New agent platforms can build their own network effects. More users bring richer feedback; more merchants widen the services an agent can call; more tools allow more complex tasks; higher completion rates draw more users. That loop could make a few large agent platforms the new centers of economic coordination.

On November 10, 2026, OpenAI launched Businesses in ChatGPT as a beta: businesses can claim their profile, maintain product data (including price and availability) and, once verified, build a Business Agent to talk with prospective customers. The first Business Agents are open only to US e-commerce and online-service companies, and those agents can also serve as destinations for ChatGPT ads. The Next Web It is an early signal, but it is still mostly about discovery and consultation, and does not mean complex transactions can already be completed autonomously.

A longer-term question is emerging: are we moving from an internet organized by a few large websites to an economy coordinated by a few large agent platforms? If so, the agent era is not a victory for decentralization but another concentration of digital intermediary power.

6. When search stops being scarce, credibility may become the scarce resource

In 1970, George Akerlof explained in The Market for "Lemons" how information asymmetry leads to adverse selection: when buyers cannot tell quality apart, low-quality goods can crowd out high-quality ones.

Agents will not make that problem go away; they may give it a new form. Once merchants optimize their data, descriptions, interfaces and quotes for agents, markets will fill with commercial information designed for machines, and it will not all be reliable: inventory that looks structured but is already stale, refund limits hidden in terms that are hard to compare, even malicious content that steers an agent into actions beyond what the user authorized. When the participants shift from people to software agents, misrepresentation, fraud and strategic manipulation do not disappear; they gain the scale of automation.

So the key infrastructure of AME may not be a stronger recommendation system but a system for verifying economic facts. A reliable delegated transaction has to answer at least five questions:

1. Does the product information come from a verifiable, real supplier?
2. Does the agent executing the trade have a valid identity?
3. Did the user authorize this kind of trade, this amount and this term?
4. Does the result match the conditions committed to beforehand?
5. If execution fails, delivery goes wrong or fraud occurs, who is liable?

These are not purely problems of model reasoning. They involve cryptography, identity management, contract law, payment networks, risk control, audit and legal remedy. The Machine Payments Protocol that Stripe and Tempo launched in March 2026 (Stripe), and the Universal Commerce Protocol launched by Google with Shopify and other retailers, are early attempts at transaction infrastructure that machines can call. But a payment standard solves only part of the technical problem of a transaction. It cannot by itself settle disputes over quality, contractual liability, truthfulness of information or platform conflicts of interest.

That suggests a valuable business category may emerge: institutions that provide verification and coordinate liability for the agent economy. They may not serve ordinary consumers directly but agent developers, merchants, payment networks and other platforms. What they sell is not information itself but the ability for information to be trusted, actions to be verified and transactions to be held to account.

7. Business models are shifting from the right to use software toward economic outcomes

For more than a decade SaaS has been the software industry's most important business model: companies buy accounts, seats and subscriptions to get tools that make staff more productive. It rests on one premise: people operate the software, and the software is a tool for people to do their work.

When agents can execute workflows directly, that relationship starts to change. A company processing 100,000 invoices a month traditionally buys finance software and hires people to read, enter, approve and reconcile them. An agent-native system might take the invoices directly, extract the data, check the rules, flag exceptions and post the entries, handing only high-risk items to a person. What the company cares about is no longer how many accounts it holds but how many valid invoices were processed, the error rate and how much human intervention exceptions needed. The unit by which software is valued may shift from the right to use it toward the result of the task.

That does not mean subscriptions disappear. More likely several models coexist: base capability by subscription, compute and tool calls by usage, clearly verifiable deliverables by outcome, and some transaction services by deal value or value realized.

The deeper effect is that when software can do work that used to need people, the line between software companies and service companies blurs: a company that appears to sell software may really be doing what outsourcers, agencies or professional service teams used to do. Outcome pricing has its own constraints, though. Results have to be attributable, both sides have to agree on what counts as success, and liability for exceptions has to be divisible. Otherwise a company has only swapped arguments over seat pricing for arguments over how outcomes are measured. The important innovation in the agent economy is not only falling technical cost but whether new service contracts, performance measures and ways of allocating liability can be built.

8. Will agents change the boundaries of the firm?

Push Coase a step further and a more interesting question appears: why do firms employ large numbers of people instead of buying each service on the market? One reason is that under some conditions internal management and long-term contracts are cheaper than frequent market transactions.

If agents can keep searching for suppliers, coordinate tasks, verify results and handle settlement, the coordination cost of using the outside market falls. A small team may get some of the operating capacity of a large company without building every internal function. A cross-border e-commerce business, for instance, could use several agents to coordinate advertising, procurement, logistics, customer support and financial reconciliation.

That does not mean firms will be replaced by agents. Firms exist not only to lower communication costs; they also make asset investments, carry organizational learning, share risk, coordinate labour and make long-term strategic decisions. When work involves highly specific assets, complex tacit knowledge or results that are hard to verify, internal organization may still beat buying from the market.

So the effect of agents on organizational boundaries will likely vary by industry. Where work is standardized, decomposable and easy to verify, outside procurement will grow; where it needs long-term collaboration, deep professional judgement and shared uncertainty, internal organization keeps its edge. The opposite may even happen: large companies, with more data, compute and proprietary workflows, may make better use of agents than small ones. The outcome depends on how technical, organizational and institutional costs change relative to one another, not on agent capability alone. That is why the agent era cannot simply be equated with an era of one-person companies.

9. Who captures the value the agent economy creates?

Even if agents lower transaction costs across society, it does not follow that everyone benefits equally. Efficiency gains and the distribution of those gains are two separate questions.

When a consumer finds a cheaper product through an agent, the direct effect is a larger consumer surplus. But if an agent platform controls most of the demand entry points, it can charge merchants distribution fees; merchants unable to reach consumers directly come to depend on a few agent platforms; and the companies supplying the underlying models, payment capacity, real supply and identity verification will all fight for a new share of the value.

That leaves at least three possible market structures:

Given the distribution advantages of existing platforms and the particular compliance and fulfilment demands of finance, medicine and logistics, I think the hybrid structure is the mid-term scenario most worth studying. That is a conditional judgement, not an inevitable technical path.

What deserves most attention is that agent efficiency and platform market concentration can rise at the same time. If an agent platform lowers users' search costs and then charges higher intermediary fees through commercial ranking, part of the efficiency gain flows to the platform rather than to consumers. So the agent economy cannot be judged only on task completion rates or transaction volume; you also have to watch consumer surplus, supplier margins, barriers to entry, platform take rates and data portability. The goal is not a market with more agents but a market that remains genuinely competitive after it has more agents.

10. From economics to sociology: when choosing itself is outsourced

Economics cares about allocating resources efficiently, but human life is not only an allocation problem. Consumption is also social behaviour: people express identity through what they buy, explore the world through travel, build relationships through where they choose to eat, and form knowledge by researching and comparing for themselves. Some of this looks inefficient but carries social and cultural meaning.

Suppose your agent could make almost all of your everyday consumption decisions, knowing your budget and your taste and always finding better value. In efficiency terms that is progress. But if you hardly browse any more, rarely stumble on a new brand and no longer deal with unfamiliar merchants, the way you encounter the world changes too. The traditional internet decides what you can see through search and recommendation; an agent goes further and decides which options deserve to be shown to you at all, and which things never need your personal attention. That is a deeper mediation of choice.

It could lighten the cognitive load of daily life and free time for creating, learning and relationships. It could also make people increasingly dependent on a few systems to understand markets and organize their lives for them. More worrying, an agent that optimizes on past preferences for long enough tends to repeat the choices you have already shown rather than encourage exploration. If someone's spending, reading, travel and learning are all coordinated by the same optimizing system, will there be fewer happy accidents in their life?

Recommendation accuracy cannot answer that. Human value does not come entirely from reaching goals; exploring, trying and failing, hesitating and changing one's mind are themselves how preferences and character develop. A good agent should not necessarily make every decision for you. It should also know which decisions deserve to stay with a person, when to raise a dissenting view and when to put new information in front of the user. Its success should be measured not only by fewer steps but by whether it extends human autonomy or gradually replaces it.

11. The reality in 2026: how far are we from this economy?

Back to reality. As of November 2026, AME is still at an early stage.

The publicly verifiable agent traffic data have a very consistent shape. HUMAN Security found that across 2025, traffic from AI agents and agentic browsers grew 7,851% year over year, yet 77% of it landed on product and search pages and only 2.3% on checkout pages (HUMAN). In its April 2026 monthly report, checkout and payment made up 3.16% of agentic traffic (HUMAN). These are monitoring samples from a single security vendor and do not represent the whole internet, and reaching a payment endpoint is not the same as completing a purchase. But they support a cautious judgement: agents have begun to change how information is discovered, while autonomous transactions lag clearly behind information retrieval.

That matches the logic of how technology develops. A failed search usually costs only time; a bad recommendation skews a choice; an execution error in a transaction causes direct economic loss. As tasks move from processing information to acting in the real world, the risk and liability a system carries rise quickly.

So over the next few years, whether the agent economy is maturing should not be judged by agent traffic, token consumption or model rankings, but by these:

• Can agents reliably complete real transactions within clearly authorized limits?
• When a transaction fails, can it be recovered or properly compensated?
• Can agents on different platforms interoperate?
• Can consumers control and move their own agent identity and preferences?
• Do merchants get a fair chance to compete?
• Do the efficiency gains agents create eventually turn into lower prices, better service and broader market participation?

These questions separate technical demos, commercial growth and real institutional change.

12. The real opportunity is not more agents but rebuilding the markets agents can take part in

For founders, the easiest mistake is to read the agent economy as a new software category and reach straight for shopping agents, travel agents, finance agents and recruiting agents. Those may well be valuable. From an industry-structure view, though, another kind of opportunity may last longer: as more software agents take part in real economic activity, the market needs infrastructure that lets them work together safely.

Suppliers need to provide real, timely, verifiable information about goods and services; users need to set authorization scope, budget limits and risk preferences; systems need execution that is idempotent, consistent and recoverable; payment providers need to tell legitimate machine transactions from fraud; and counterparties need to record commitments, verify fulfilment and assign liability when disputes arise. None of this goes away as base models get stronger. Better models lower reasoning costs and some operational errors, but they cannot replace property rights, contractual relationships and liability in the real world.

Working from first principles, I think the most valuable companies in the agent economy will compete on three layers:

The three layers compete and also cooperate. The question a founder really has to answer is not "can I build an agent" but: when an ever smarter agent tries to complete some real economic activity, what is left that it cannot solve by reasoning alone? Wherever that answer sits is usually where the opportunity worth studying is.

Conclusion: the next internet may be a system for economic coordination

Important technological revolutions usually change not only which tools people use but how society organizes economic activity. The internet let information cross geographic borders, the mobile internet made transaction entry points available at any moment, and cloud computing lowered the cost of getting compute. Agents may go further and lower the coordination cost of turning human intent into action across systems.

AME does not mean machines become the end purpose of the economy, nor that consumers, firms and traditional markets are replaced by automated systems. People still define needs, bear responsibility and enjoy what goods and services give them. What changes is how people take part in economic life: we used to enter markets ourselves to search, compare, negotiate and close; in future we will increasingly take part through our own digital agents. Commercial competition will extend from fighting for human attention to fighting for agents' choices and calls.

It also brings back some old questions: who holds the information, who has the decision rights, who controls the entry point to trade, who bears the risk, and who captures the value that growth creates? The questions economists from Coase and Hayek to modern platform economics have argued over will not vanish because AI has arrived. As more economic decisions are handed to software, they only matter more.

B2A describes how businesses face a new kind of software customer; AME asks what happens to the market itself when those software customers start coordinating economic activity on people's behalf. The first is an extension of a business model; the second may be a change in how the economy is organized. The next internet need not be an even larger network of content. It may gradually become an economic system in which people set the goals, agents coordinate the tasks, firms provide real supply and trusted institutions guarantee the transactions.

Whether that brings more open markets or new platform monopolies has no answer yet. Technology can lower coordination costs, but it cannot decide on its own who captures the gain. So what really deserves attention is not just how much agents can do for people, but whether, as more decisions stop being carried out by people themselves, we can build a new economic order that raises efficiency while keeping market competition, individual autonomy and accountable institutions intact.

The next internet revolution may not be about giving everyone a smarter assistant, but about the whole market learning to run alongside those assistants. What decides its meaning in the end is not how many times machines can trade, but whether people actually come away with more freedom and more choices.

References

• R. H. Coase, The Nature of the Firm, Economica, 1937.
• F. A. Hayek, The Use of Knowledge in Society, American Economic Review, 1945.
• G. A. Akerlof, The Market for "Lemons": Quality Uncertainty and the Market Mechanism, Quarterly Journal of Economics, 1970.
• J.-C. Rochet & J. Tirole, Platform Competition in Two-Sided Markets, Journal of the European Economic Association, 2003.
• OpenAI lets businesses claim ChatGPT profiles and launch their own agents, The Next Web, 2026-11.
• Introducing the Machine Payments Protocol, Stripe, 2026-03.
• The 2026 State of AI Traffic & Cyberthreat Benchmark Report, HUMAN Security.
• State of Agentic Traffic – April 2026, HUMAN Security.

Personal notes, not investment advice.