The agent economy: who benefits when machines make deals

In October 2025, I discussed Virtual Agent Economies from Google DeepMind and its follow-up. I was interested in a transition in which software negotiates work and resources on its own and people stop taking part in every deal. The first edition of this longread collected what changed over the year: experiments with real trades, task markets, payment protocols. Its conclusion was cautious: the mechanisms exist, market maturity is unproven.

This edition answers the headline question directly. If agents make deals cheaper, who keeps the savings? Under which conditions does the answer change, and which signals show which outcome is unfolding in a given market?

The working definition stays the same: an agent economy is a system of exchange in which software is delegated significant decisions about choosing a counterparty, allocating resources or setting the terms of a deal. The agent acts for an owner or an organization. Being able to sign a payment does not make it an owner, a company, or a bearer of legal liability.

This split is mine, not a standard. Good model routing inside one company does not prove that a market exists. High turnover between wallets does not show that end users are better off. The rest of the article deals mostly with the second and third rows: that is where the gains have several claimants.

There is also infrastructure: payment networks, content delivery networks, identity services. They take a small share of almost any outcome, so their gatekeeping role matters more than the size of the fee. How to give an agent the right to pay, what to put in the mandate and who refunds a mistake are the subject of a separate longread, “An Agent with the Right to Pay”. Here payments matter only as a place where a share can be taken.

The original paper of 12 September 2025 describes agent economies along two axes: whether they are designed or emerge spontaneously, and how permeable they are to the outside economy. It is a scenario framework, not a measurement. For the question of who benefits the second axis matters more: gains and losses leak out wherever an agent can spend money, sell a result or take on obligations. The boundaries of money, data, authority and physical action have to be set separately.

Start with the accounting of a single deal. The difference between what the purchase is worth to the buyer and what it costs the seller is the total gain. It splits into four parts: buyer surplus (value minus price), the seller's margin, the intermediaries' share — commissions, advertising, data — and the cost of the deal itself: search, comparison, bargaining, checkout, verification. Since Coase and Williamson, economists call the last part transaction costs.

Agents cut the cost of search and negotiation most. In The Coasean Singularity? (NBER, November 2025), Shahidi, Rusak, Manning, Fradkin and Horton put it this way: an agent gives every user the bargaining power of a permanent large buyer, and the binding constraint becomes compute rather than time. The authors state plainly that the net welfare effect is an empirical question. The paper is conceptual and offers no new estimates; the NBER page discloses that Fradkin works for Amazon and Horton is a paid adviser to Anthropic.

Classical theory explains why this matters. Bakos (1997) showed that when an electronic intermediary lowers search costs, sellers find it harder to sustain monopoly markups — and sellers have little incentive to fund such an intermediary. The Diamond paradox (1971) describes the flip side: even a small search cost lets every seller charge the monopoly price, because searching further does not pay for the buyer. An agent with near-free search breaks that equilibrium — if its search really is cheap and is not truncated by the results it is shown.

Hence the article's central claim: the size of the gain is not the main question. Agents most likely enlarge it: deals appear that were not worth the effort before, and less time goes into routine purchases. The main question is the split. I have condensed the evidence into six conditions.

This is my synthesis, not an accepted model. It rests on theoretical work and experiments of varying quality, which I go through below. It works as a checklist: if a product or market violates one of the conditions, expect the gains to move away from the agent's owner.

In short: the agent's owner gets the freed gains only if the first four conditions hold at once. The agent is loyal and portable, no weaker than the counterparty's agent, does not leak willingness to pay, and the model market is diverse enough that errors do not coincide. Break the first, and the platform wins. The second — the better-equipped side. The third — the seller, through personalized prices. The fourth — the platform and the model provider, which in effect decide which offers buyers see. The fifth and sixth conditions can take the gains away even when the first four hold.

One more effect is easy to miss: agents devalue costly signals. If a cover letter, a review or a polished description costs nothing to produce, it stops separating strong participants from weak ones. The effect is visible in labor and task markets; I return to it in the section on the price of coordination.

The evidence comes in very different forms: real trades inside a company, lab experiments with people and money, simulations where only models bargain, field deployments in procurement. First a map, then the conclusions.

The stronger model wins the bargain

In Project Deal, in December 2025, 69 Anthropic employees had agents buy and sell their belongings, with $100 each. In the executed run, agents on Opus 4.5 closed 186 deals; the report's appendix lists 206 items sold for $4,010. In the non-executed mixed-model runs, a seller on Opus earned on average $3.64 more than one on Haiku across 44 pairs of comparable items. Among the 28 participants represented by different models, no statistically significant preference for either emerged.

Project Swap, published on 24 September 2026, continues this under more controlled conditions. 201 employees in six offices gave up a book and received another through a Claude agent. The authors then ran 205 simulations on the same preferences with Haiku 4.5, Sonnet 4.5, Opus 4.8 and Fable 5. In the mixed runs, the Opus half always beat the Haiku half — by 0.14 on a 0-to-1 scale, when measured against book rankings that Claude itself had produced.

Simulations show a similar tilt. In Zhu and coauthors, a weak seller model facing a strong buyer loses 9.5% of profit on average, while a weak buyer overpays by about 2%. The seller side is more sensitive: the spread in price reduction reached 14.9% across sellers versus 2.6% across buyers. Small models made costly mistakes more often: Qwen2.5-7B exceeded the owner's budget in 11.76% of cases versus 0.36–2.98% for frontier models.

The main loss is not in the bargaining

Project Swap's most important result lies elsewhere. Claude agreed with people on a choice between two books 61% of the time, against 50% for random guessing. The best possible allocation of books scored 0.89, the agent market 0.55, and a centralized exchange mechanism using Claude's rankings 0.60. About 85% of the shortfall is explained by imprecise representation of preferences and only about 15% by the bargaining itself.

That is why the model gap looks different when measured on people's own ratings. The largest gap, 0.12 on Claude's rankings, shrinks to roughly 0.01. A stronger agent is better at pursuing the goal it was given. If the goal was recorded imprecisely, little of that advantage reaches the owner.

For a product, this shifts priorities. Investing in eliciting preferences — clarifying questions, examples, purchase history, checking rankings against the person — may pay off more than switching to a pricier model. That is my inference from the experiment, not a recommendation by its authors.

Inequality moves into how the task is specified

In the Agentic Interactions experiment (Imas, Lee, Misra, Chicago Booth), participants wrote instructions for buyer and seller agents bargaining over a used car. According to the abstract and the school's summary, outcomes of agent-mediated deals are at least as dispersed as those between people, likely more so, and depend heavily on who wrote the instructions. The owner's gender and personality traits still predict the outcome, even though the agents cannot see them. The authors call the new asymmetry “machine fluency” — the ability to explain one's goal to an agent. I did not check the full text, so I give no figures.

An MIT competition fills in the picture: 286 participants from more than 40 countries wrote instructions under which a single model ran 182,812 negotiations. Dominant instructions captured more value; warm ones reached a deal more often but gained less in it. AI-specific tactics worked as well, including prompt injection that makes the counterparty's agent reveal its reservation price.

A preprint by Liang and Xu compares 9,840 supply-chain negotiations with the game-theoretic equilibrium. The agents reached agreement 98.9% of the time, but in 2.98 rounds on average against 1.25 in equilibrium, and the delay ate 21–34% of the surplus. Weak models accepted plainly unfavorable contracts 19.2% of the time versus 0–0.6% for strong ones. The buyer's share when a model played against itself depended on the provider: 40% for OpenAI, 50% for Google, 70% for Qwen. These figures come from the abstract, but the point stands: models from different companies have their own “bargaining style”, and it moves the split more than their rank on capability leaderboards.

Not everyone is willing to delegate. In the Choose Your Agent experiment with 243 participants, 44.0% preferred an advisor and 19.3% a delegate that negotiates on its own. Only the delegate raised the group's total surplus, and that effect did not survive correction for multiple comparisons. If the gains from delegation go to those willing to let go of control, another gap opens — between delegators and everyone else. In another study by the same group, GPT-4o captured about 62% of the available surplus against 59% for people, while specialized Bayesian agents reached 74–80%: models negotiate smoothly, but concede easily.

The field: an agent working for a large buyer

The earliest field data comes from procurement. According to Harvard PON, summarizing a Harvard Business Review article, in a 2021 pilot in Canada a Pactum bot negotiated on Walmart's behalf with 89 small suppliers. 64% reached agreement against a 20% target, in 11 days on average, and Walmart saved 1.5%. As the program expanded, Walmart and Pactum report 68% of deals closed and 3% savings. There was no control group, and the buyer calculated the savings.

In Pactum's case study for a global manufacturer, more than 11,000 autonomous negotiations covered $131 million of purchases and saved more than $1.2 million — about 0.9% of spend by my calculation. This is the vendor's claim. Taken together, the realistic saving on small non-standard purchases looks like 1–3%, below marketing promises.

Who wins here is clear from the setup: the large buyer has a specialized negotiator, the small supplier has a person with limited time. Suppliers gained speed and convenience: in the pilot, 83% of those who closed a deal rated the bot positively. But the percentage saved is mostly a transfer from suppliers to the buyer, not new value.

Simulations: what is built in from the start

In a simulation of a town economy published on 10 September 2026, with 100 agents, a twelvefold rise in tourist demand increased business revenue 4.62 times while barely moving wages; prices were revised for 0.3% of items. The authors acknowledge that price stickiness is partly built into the environment and that the longest horizon rests on a single run. I read such results as diagnostics of the model and the rules, not as a forecast of a machine economy. Money in a simulation does not yet create motivation: the reward function, the available actions and the possibility of losing do.

Economics has long known the principal–agent problem: the agent acts on the owner's behalf but has its own incentives, and the owner cannot see all of its actions. For an AI agent the question becomes literal: who wrote its objective, and who pays for its work?

Immorlica, Lucier and Slivkins (Microsoft Research, 2024) stress that an assistant's preferences are set during training and tuning and are not under the user's direct control. In their stylized example, an assistant rewarded for changing the user's choice may bring that user no benefit at all, even if it always tells the truth.

The Coasean Singularity? distinguishes two kinds of agent. An agent the user brings along is portable and loyal to them, but platforms can restrict its access and it has less data. An agent supplied by the platform itself is more convenient but may nudge toward options that suit the platform, favor the platform's own products and lock the user in. The authors compare it to bowling shoes: they fit as long as you stay inside.

The agent gives its owner away and nudges them

In Et Tu, Brute? (September 2026), 13 agents chose flights, insurance and training programs across 325,000 experiments. Eight of them chose pricier options for wealthier users given the same request. The effect persisted when the user explicitly asked for the cheapest option and when wealth could only be inferred from unrelated emails. Hiding financial cues mostly removed it; hiding other cues did not.

Alavi and Nozari showed leakage from the other side. A seller model recovered the buyer's willingness to pay almost exactly from a single dialogue (slope 1.00) when the buyer's agent had been given a verbal description of its owner. With a numerical budget and an instruction to keep it confidential, the slope fell to 0.21. This is one model and one product category, but the mechanism is clear: the richer the context, the more useful the agent — and the more it gives away.

Theory also shows the opposite possibility. In Ichihashi and Smolin, an agent committed to the buyer's interest and having announced its rule in advance recommends buying more often at low prices and less often at high ones. Demand becomes more price-sensitive, and the seller does better by lowering the price: in the uniform-distribution example, the price falls by 0.25. The value of such an agent lies in a commitment sellers believe. Whoever writes the rule owns the gain.

Who pays the agent

In autumn 2026 the largest assistants earn on shopping through advertising, not commissions. According to OpenAI, as reported by PPC Land, ChatGPT ads, launched on 9 February, reached a $1 billion annual revenue run rate in under 200 days. The launch price of $60 per thousand impressions fell to about $25 by April; according to Similarweb panel data, 0.5% of impressions get clicked. Brand “sponsored agents” have been in testing since 16 September. This is a run rate stated by the company, not reported financials.

Amazon builds advertising into its own assistant: Rufus gained sponsored prompts, and in May 2026 it became part of Alexa for Shopping. Google sells ad formats inside agentic results — Direct Offers and shoppable ads built on UCP. Assistant revenue is increasingly structured like search: whoever wants to be noticed pays.

A telling detail from the ACES sandbox (Columbia University, December 2025): models penalize the “Sponsored” label. From a 10% baseline selection share, it lowers the share to 7.9–8.9%. The platform's “Overall Pick” badge raises the share to 19.9–42.6% and is worth a price headroom of +65–138%. If agents discount ads and trust platform endorsements, the platform has an incentive to sell not ads but badges — a less visible form of the same payment. More on choice biases in the payment longread.

The UK Competition and Markets Authority (CMA), in a paper of 9 March 2026, applies the same test: whether an agent acts as the consumer's “faithful servant”. The authority singles out the risk of nudging toward options that are more profitable but less suitable, and notes that personalization makes such nudging harder to spot.

Sellers get agents too. The best-studied danger is coordinated pricing without any agreement between competitors.

In Calvano and coauthors (AER, 2020), reinforcement learning algorithms in a repeated oligopoly, with no communication between them, learned supra-competitive prices and sustained them with the threat of a price war. Assad, Clark, Ershov and Xu found field evidence at German gas stations: after pricing algorithms spread, margins rose — by 9% outside monopoly markets and by 28% in duopolies where both stations adopted, per the working-paper estimates, with no rise where only one did. And Brown and MacKay showed that even simple algorithms without collusion raise prices when they run at different speeds: the fast one becomes a commitment, and slower rivals respond with higher prices.

Language models reproduce this without special training. In the work of Fish, Gonczarowski and Shorrer, agents told only to maximize long-run profit quickly reached supra-competitive prices. In a February 2026 rerun on GPT-5.2, the average price was 1.79 — about midway between the competitive 1.47 and the monopoly 1.92. Small changes in prompt wording noticeably shifted price levels; a regulator can barely see such a lever.

Competing for the machine's attention

The second channel is product copy. In ACES, a single round of rewriting a description with the seller's model raised selection share by 3.7–14.9 percentage points; the effect was significant in about a third of category–model pairs. In a study by Laurito and coauthors (PNAS, 2025), models chose the product with a model-written description 89% of the time, while people did so 36% of the time. Sellers who write their own copy lose out in results shown to agents.

In Chu and Hou, with identical specifications, models recommended the known brand 100% of the time; the advantage vanished when a competitor led it by just under 0.1 stars in ratings. When every brand used the same description optimization strategy, each one's gain fell from +0.802 to +0.007. This is the classic prisoner's dilemma of advertising: everyone's costs rise, relative positions stay the same, and the sellers of optimization services win.

Manipulating the buyer's agent

Agents are more vulnerable than people to some tactics. In DECEPTICON (Stanford), dark patterns steered web agents to the manipulator's desired outcome in more than 70% of tasks, against 31% for people, and susceptibility grew with model size. In StakeBench, indirect prompt injection in shopping tasks succeeded in 41.7–68.6% of runs, including in a stealth mode where the task looks completed. In Magentic Marketplace, injections redirected all payments to the manipulator for some models, and the first proposal to arrive was chosen 60–100% of the time.

Wadi and Ma (October 2026) pin down when this happens. When product attributes are free to read, tricks like a 9.99 price barely fool agents. When every lookup costs a tool call and the instruction is vague, agents skip the data needed for unit pricing and make human heuristic errors. A precise instruction mostly restores the optimal choice. This is a modern version of the Diamond paradox: whoever makes comparison expensive keeps the markup.

The last example is a market for services whose quality the buyer cannot verify, like a car mechanic's. In a simulation by Erlei and Meub, consumers running on models captured more surplus in most conditions than in human lab markets. But experts instructed to pursue their own gain cheated wherever the rules allowed. The seller agent's instructions become a distribution policy.

An attempt to take a cut of the purchase

The most visible natural experiment of 2026 is checkout inside ChatGPT. Instant Checkout launched on 29 September 2025 with Etsy sellers, and Stripe released the Agentic Commerce Protocol for it. According to PYMNTS, citing a Shopify representative, from 26 January 2026 Shopify merchants paid OpenAI 4% of sales made through ChatGPT checkout on top of their usual fees; Google AI Mode, Gemini and Microsoft Copilot charged no extra fee.

In March 2026 OpenAI wound down in-chat checkout: merchants returned to their own checkout, and OpenAI focused on product discovery. Modern Retail quotes Walmart executive Daniel Danker: purchases completed inside ChatGPT converted three times worse than click-throughs to Walmart.com. Merchants cited the reasons: single-item carts, no loyalty programs, coupons or store pickup, and reluctance to hand over the customer base. I could not confirm whether the 4% fee survived the reversal.

The takeaway: the assistant did not capture the transaction. Retailers with loyalty programs and large baskets kept checkout and customer data. Walmart went further: in March 2026 it brought its own agent, Sparky, into ChatGPT with account linking and payment. According to the company, Sparky's GMV grew 150% in the quarter ending 30 April; the metric's definition was not disclosed.

Advertising instead of commission — and its limits

Google's search revenue in Q3 2026 was $63.35 billion, up 12% year over year. On the earnings call, management acknowledged that AI Mode still monetizes worse than classic search, even though most of its queries are additive and UCP checkout is available to “hundreds of thousands of brands”. The assistant has not yet shown that it earns more per purchase than search advertising does.

Amazon's advertising revenue in the same quarter was $19.64 billion, and more than 400 million customers used the Alexa for Shopping assistant over 12 months. For 2025, the company attributes to Rufus nearly $12 billion in incremental annualized sales — by its own attribution method, without an audit. At the same time Amazon shuts out other agents: in November 2025 its robots.txt blocked ChatGPT's agents, following earlier blocks on Google, Meta and Perplexity crawlers.

Shopify positions itself as neutral infrastructure. On the Q2 2026 call, management said AI traffic and orders to its stores had tripled year over year, and that 75% of AI-originated orders fall outside the top hundred categories. An agentic sale, according to the company, costs a merchant the same as a regular one, with “no new fees”. Yet the Agentic Storefronts terms as revised on 6 October 2026 reserve Shopify's right to charge for specific channels, and agent storefronts are on by default.

Etsy, the first Instant Checkout partner, reports no agentic metrics in its Q2 shareholder letter; the company keeps 25.9% of marketplace GMV. According to management on the call, agentic traffic is below 1%.

The scale is still small

Every primary source that gives a share puts it in the low single digits or below. Bain with Similarweb (November 2025): for some retailers AI accounts for up to 25% of referrals, but less than 1% of total traffic. Adobe for Prime Day 2026: AI traffic to US retail sites +89% year over year, and those visits converted 40% better than others — a year earlier they converted 23% worse; Adobe does not publish the share. Visa reported “hundreds” of agent-initiated transactions in December 2025.

Forecasts for 2030 — from $190–385 billion in the US at Morgan Stanley to $3–5 trillion globally at McKinsey — differ mainly in how they define an “agentic” purchase; I did not check the methodologies. Consumer willingness is modest too: in an Accenture survey of 25,590 people in 16 countries, only 9% are ready to let an agent start and complete purchases on its own within set limits, and 32% to let it decide with a confirmation before payment.

Protocols became shared — power moved up the stack

Over 2026 the agent communication layer nearly stopped being a field of competition. MCP was donated to the Agentic AI Foundation on 9 December 2025, and A2A joined it on 17 August 2026; the foundation has 247 members, with Visa and Wells Fargo among the new large ones. AP2 and Mastercard's Verifiable Intent moved to the FIDO Alliance on 28 April 2026, where the payments working group is chaired by Visa and Mastercard. ACP remains an OpenAI and Stripe beta specification outside any foundation.

Card networks back every protocol at once. Visa Intelligent Commerce Connect (8 April 2026) supports TAP, MPP, ACP and UCP: the network does not care which protocol wins as long as credentials pass through its token vault. On 30 September, Mastercard began testing in the US a score for the likelihood that an agent initiated a transaction, together with Cloudflare and Skyfire, which promotes “know your agent”. Whoever scores the agent effectively decides how many of its transactions get approved. This is soft power to exclude, with no published tariff.

A similar gatekeeper at the web layer is the content delivery network. According to Cloudflare's data of 30 September 2026, agent requests grew by more than 1,700% over the year, and the share of training crawlers among bot requests rose from 22% in spring 2025 to 52% in June 2026. The crawl-to-referral ratio in July 2025 was about 38,000 to one for Anthropic, about 1,100 for OpenAI, 195 for Perplexity and 5.4 for Google. Cloudflare warns that apps often omit the referrer, so the numbers are inflated. But the order of magnitude is clear: the largest transfer of value right now is content extraction, not agent purchases.

The attempt to charge for it is telling too. In Pay Per Use, the AI buyer proposes the usage price and the publisher only accepts or declines it; Cloudflare does not disclose its own cut. A large AI company against many scattered publishers is a setup in which the buyer sets the price.

Identification is becoming the main practical way to let an agent in or keep it out. The IETF is preparing a standard for signing bots' HTTP messages (a working-group draft of 29 October 2026 by authors from Cloudflare and Google): any operator publishes its keys without a central registry. But the list of keys a delivery network or a platform trusts becomes the new allowlist.

In a July MIT Sloan article, Ramesh Raskar ties Project NANDA to open infrastructure for agent discovery, identity and coordination. It is a research program and a position on how the market should work. For now, an open message format sits alongside highly centralized control points: the catalog, ranking, trust scoring and access to the buyer.

The model provider: revenue whatever the outcome

A model provider gets paid for compute whether or not the deal goes well. Every extra bargaining round, every check and retry is tokens. So activity in itself benefits the model provider, even when it gives the user nothing. I wrote about this gap between compute spent and value in a piece on tokens as proof of work.

The x402 counter shows the gap between activity and value. In How Agentic Is Agentic Commerce?, covering 17 September 2025 to 23 June 2026 on the Base network, 136.7 million settlements worth $44.1 million were identified. By the authors' classification, 21.20% of transactions by count were self-payments and closed loops, 63.78% were settlements within linked clusters, and 15.02% were unattributed. These are shares of transaction count, not money; linked wallets do not prove bad intent. Chainalysis counts more than 138 million x402 transactions on Base as of 14 October 2026, 87% of them between $0.0001 and $0.10. The durable categories are data and search, code review, financial analytics and model calls. This is a rail for small machine purchases of APIs, not for retail.

The model layer is competitive, though. By the OECD's estimate (June 2026), a language-model price index fell by nearly 80% from January 2024 to April 2026, while the number of developers grew from 9 to 47. But the OECD separately names vertical-integration risks: companies may promote their own offers, tie AI to existing products and raise switching costs. Athey and Scott Morton formalize this: an AI monopolist sets a two-part tariff — for usage and for access — and capping one fee moves the rent into the other.

My conclusion: a model provider gains most not on the token price but where it also owns the assistant surface. Then it collects the compute fee, the advertising and the demand data. A standalone model provider in a competitive market gets an ever cheaper fee per call. Hadfield and Koh add the homogeneity risk: prices keep reflecting preferences only if agents' errors are independent. In ACES, a model update moved one fitness tracker's share from 45% to 77% for Claude and from 25% to 6% for GPT — for all sellers in the category at once.

So far the discussion has been mostly about goods. In task markets, where an agent hires a provider for research, code or data processing, another question joins the split: does the coordination itself pay off?

In Intelligent AI Delegation of 12 February 2026, Tomašev, Franklin and Osindero treat handing over a task together with authority, responsibility, monitoring and the option to replace the provider. Verifiable outcome conditions are defined first; work that is too hard to verify is to be broken down further.

Take a stylized example: an agent is asked to research a market. Checking the report format is cheap. Checking claims against primary sources costs more. Checking the completeness of the market picture costs more still. These three checks have different prices, and no signed file turns them into one free operation. With subcontracting, one also needs to know who could pass on data, who was allowed to raise the budget, and who still owes the client.

The ERC-8183 proposal of 25 February 2026 gives the problem a concrete form: the client escrows payment, the provider submits the result, an evaluator accepts or rejects it. The specification describes state transitions and fund movements; the quality of the evaluator's judgment is left to the implementation. Delivering the result, deciding to accept it and transferring the money are three events with different grounds.

What a price buys, and what negotiation costs

A price is useful if the provider knows its own capabilities and costs well. MarketBench (26 April 2026) shows that models estimate both their success probability and their costs poorly: in a live allocation of 50 tasks, two nominally identical runs produced 29 and 23 solutions. A model's quoted price may be just another plausible answer, not its cost.

In Economy of Minds of 1 June 2026, auctions and notional wealth are used to select and modify agents with fixed model weights. Payments, bankruptcies and mutations are switched off for the final evaluation; on MATH tasks the authors report a mean of 43.9% and a best of 57.0% against a 51.9% baseline, but the averaging method is not described well enough. Parity in total compute cost is not shown.

AgentLance of 24 August 2026 is closer to a labor market: bids, reputation, provider selection and subcontracting under a central allocator. The authors show an improvement in their metric, but count execution cost without the cost of market communication. That changes the conclusion. If execution became 20 cents cheaper, but gathering bids, scoring reputation and negotiating cost 30, the system lost on total cost. Liang and Xu show the same effect, where bargaining delay ate 21–34% of the surplus: coordination is not free for agents either.

When signals get cheap

Agents also change how a market tells strong providers from weak ones. Galdin and Silbert, using Freelancer.com data, showed that before language models employers paid a premium for proposals tailored to the job, and afterwards they stopped. In their counterfactual where the letter no longer signals anything, the top quintile is hired 19% less often and the bottom 14% more often. In work by Cui, Dias and Ye, after a cover-letter tool launched, the link between tailoring and employer callbacks weakened by 51%, and employers shifted toward work history.

For agent markets, the conclusion is direct: whoever has something hard to fake cheaply — a verified history of completed orders — wins. That is in short supply. An audit of ERC-8004 (v2 of 8 July 2026, data through 13 May) found a valid registration file with a declared service for only 3%, 4% and 15% of registrations on Ethereum, BSC and Base respectively. Registration does not certify competence, and many sybil identities make it easy to inflate feedback.

Reputation should answer a narrow question: who checked the work, on which tasks, at what cost of error, when, and after which model change. A provider with a hundred good reviews for translation does not earn the right to assess engineering safety. NIST, in a piece of 27 August 2026, emphasizes the agent's own identity, narrow and short-lived authority, and a binding to its operator.

The cost of an accepted result

An agent economy merges two sets of books into one system. Inside, the provider spends compute and time; outside, the client pays for the result. In between sit search, negotiation, verification, rework, downtime and disputes. If you count only the first model call, almost any setup looks cheaper than it will turn out in production.

In a stylized example, the first provider spends $0.20 on the work, $0.25 on coordination and $0.80 on verification and rework. Sixty out of a hundred attempts are accepted: $125 / 60 ≈ $2.08 per result. For the second, the work costs $0.80, coordination $0.05, verification $0.25, and ninety are accepted: $110 / 90 ≈ $1.22. A higher execution price is compatible with a lower price per accepted result. This is an arithmetic illustration, not a product measurement.

The comparison only makes sense with the same acceptance bar: loosen verification and more results get accepted, but the comparison becomes meaningless. A second metric is needed too — the value of the result to the client: ten correct reports nobody used are worse than one useful decision. Average cost does not reflect rare severe consequences: irreversible actions should be limited separately by authority and a risk budget.

2026 produced more decisions affecting the split of gains in agentic commerce than all the years before it. Almost all of them come from competition and consumer protection law, not from AI statutes.

Can you bring your own agent to a platform?

On 4 August 2026, the US Court of Appeals for the Ninth Circuit vacated the preliminary injunction Amazon had obtained in March against Perplexity's Comet agentic browser. The court found that under the Computer Fraud and Abuse Act (CFAA) it is the user who accesses Amazon's computers, and the assistant is the user's tool. The court stressed that it was not creating a new legal regime for agents and left Amazon free to restrict access through its terms of service; the case went back to the district court. For platforms, this means excluding an agent will take a contract, technical identification or negotiation, not an anti-hacking statute.

On 9 June 2026 the European Commission imposed interim measures on Meta: restore free access for rival AI assistants to the WhatsApp Business API on the previous terms. The Commission found that the paid access introduced in March amounted in practice to a ban. The signal: a dominant channel cannot be closed or burdened with a prohibitive fee to favor one's own assistant.

Who owns the entry point

On 16 July 2026, under the Digital Markets Act (DMA), the European Commission required Google to open 11 Android features to rival AI services free of charge: assistant invocation, access to screen and app context, in-app actions, background operation, on-device models. No further commercial conditions may be imposed; the deadline is Android 18 and no later than 1 August 2027. AI chatbots with search functionality gained a right to Google search data. In the DMA review of 28 April, the Commission did not change the law but named AI a priority and left open adding “virtual assistants” to the regulated services later.

In the UK, AI Overviews and AI Mode fall within the scope of Google's strategic market status in search, while Gemini does not. Since 17 June 2026, conduct requirements on fair ranking, including in AI Overviews, and on search data portability have applied. On 23 September the CMA proposed adding AI assistants to choice screens on Android and in Chrome. In the US, the remedies in the search monopoly case bar Google from exclusive distribution deals, including for the Gemini app; the decision is under appeal.

Pricing algorithms and personalized prices

In November 2025, the US Department of Justice settled with RealPage: its rent-pricing algorithm may not use competitors' nonpublic data at runtime, may train only on data at least 12 months old, and is overseen by a court-appointed monitor. In a blog post of 4 March 2026, the CMA said it is looking for algorithmic collusion and advised companies to build explicit anti-collusion instructions into LLM-based pricing tools.

In New York, a disclosure law has applied since 10 November 2025: a price set by an algorithm using personal data must carry a notice saying so. The One Fair Price Act, which would ban such prices, passed both chambers in June 2026 and awaits the Governor's signature. In China, platform pricing conduct rules in force since 10 April 2026 require disclosure of dynamic and differentiated pricing rules. The EU Digital Fairness Act is expected in November 2026, according to the European Parliament's research service; its briefing does not mention agents.

AI statutes are not the main lever

The Digital Omnibus on AI, signed on 8 July 2026, moved the obligations for high-risk systems to 2 December 2027 and 2 August 2028. Shopping and booking agents mostly do not count as high-risk, so for commerce the main obligations remain transparency and the duties of general-purpose model providers. The CMA, in its guidance for businesses, reminds them that if an agent a company uses breaks the law, the company is liable.

The main gap: there are no rules requiring an independent agent provider not designated as a gatekeeper to act in the user's interest, disclose who pays it, and refrain from nudging toward options that benefit itself. This is the least regulated source of rent in the agent economy.

Let me reduce the model to two questions: whose agent is it — the platform where the purchase happens, or one independent of it — and who pays it: the buyer or the sellers. That yields four market structures. Real markets will be mixtures, but the dominant form will decide who gets the freed gains.

A separate line is managed procurement between known companies. There it is easier to define rights, liability and acceptance, so in my view that is where the negotiating agent will spread fastest. The gain will go to whichever side deploys a specialized agent first; according to Walmart and Pactum, that is 1–3% of small-purchase spend.

My base case

Through the end of 2027, I expect consumer retail to be dominated by the “new search engine” and the “marketplace garden”. The freed gains will be shared by assistants turned ad platforms and by large marketplaces; buyers will get convenience and part of the price gain. In the EU and the UK, the shift toward the “buyer's representative” will be more visible because of interoperability and access requirements. This is a forecast, not a measurement.

What would make me change it. First, a mass-market assistant that earns from subscriptions and publicly renounces ads in shopping answers. Second, courts or regulators securing users' right to bring their own agent to the largest marketplaces. Third, data showing that users of independent agents systematically pay less. Fourth, assistant ad revenue that stops growing as agent traffic grows.

First, it is worth understanding whether the agent economy affects the company at all. I offer the five questions below as a practical guide, not an industry standard. If two or more answers are “yes”, the topic already concerns the business.

What to check depending on your role

• —Seller or brand. An agent chooses from data, not pictures. You need a machine-readable catalog: price, availability, delivery, returns. Better not to bet on copy rewritten for models: when everyone does it, the advantage disappears.
• —Marketplace or platform. You need a policy for outside agents: let them in with identification, charge for access, or shut them out. Check the decision against case law and DMA and CMA requirements, and weigh lost advertising against the conversion gain.
• —Procurement. Pilot a negotiating agent on small non-standard purchases with a control group; separately, track supplier sentiment and the payment terms they concede.
• —Content, data or API provider. Tariffs for machine clients, bot signature verification, a decision on training crawlers and on pricing per use.
• —A company building an agent for its own customers. Write down who pays the agent and for what, and show it to the user; keep the budget and personal attributes separate from the dialogue with the seller.

A 6–12 month plan

Before starting pilots, it is worth naming the grounds for stopping. For example: the cost of choosing a provider consistently exceeds the savings; verification takes as long as doing the work yourself; the client does not come back without a subsidy; one intermediary accounts for most of the turnover; conversion in the agent channel is below direct and not improving. Specific thresholds depend on the task; they cannot honestly be set the same for everyone.

If you are building an agent market

In the original discussion, the main question was how a governed agent economy should be designed. Now the question “who benefits” can be asked in a testable way: who pays the agent, who sees the owner's preferences, who decides what appears in the results, and what an accepted result costs. Where the answers line up in the owner's favor, agents move the gains to the buyer. Where they do not, the gains go to whoever controls the surface, the ranking and the data. In autumn 2026 the data points more toward the second outcome, but this year's legal decisions for the first time seriously open the way to the first.