A few years from now, two purchases land in the same week. The first one you hardly notice. A bag of coffee shows up at your door that you don't remember ordering. Months back you mentioned a price to your assistant and forgot you'd said it, you were running low, and overnight a roaster hit that number, so the bag shipped before coffee had crossed your mind. No checkout, no comparison. A machine read a price off a page and decided it was good enough.
The second one is anything but quiet. Inside a mid-size factory, the woman who buys the steel comes in Monday morning and tells her agent what she needs this week and the most she'll pay. The agent works overnight. It pulls quotes from five mills, plays them against each other, and lets her current supplier match the best one. By seven there's a recommendation on her screen, most of the order going to one mill and a little to another, with a note at the bottom: prices are unusually low this week and won't hold, so maybe buy a bit more than usual.
Both are agents buying for someone. The first haggled all night; the second bought before its owner had decided anything. Why they look so different is the question.
The hunch everyone has
Ask most people what a world full of capable agents looks like and they tend to say the same thing: everything turns into a negotiation. If haggling costs nothing, why would your agent ever accept a posted price? Why not push for a few cents off the coffee and a few dollars off the flight, every time? It's an easy guess, and it's the same one people made twenty years ago, when the web was supposed to turn every market into one big open-air bazaar.
We think that's backwards. For most of what you buy, agents will kill negotiation rather than spread it, and what replaces it is closer to instant selection: your agent reads the posted prices and picks the best one. Bargaining survives, but only in a corner of the economy most people wouldn't guess, and by the time it gets there it looks more like an auction than like haggling.
Underneath sits a bigger question that has little to do with the goods themselves. Whether this leaves you better off or worse off comes down to who owns the agents. We come back to it at the end.
None of this is hypothetical. Agents already handled part of last holiday season's online orders, and the big platforms are rebuilding themselves around them. It is also our day job. We run a public benchmark that sets today's agents against each other and records what they do, so much of what follows comes from watching them up close.
The old map was drawn by a vanishing cost
Start with the map everyone carries around: the line between shopping as a consumer and buying as a business. Groceries and gadgets get a fixed price. Steel and software contracts get negotiated. We have all absorbed this as if it were a law of nature.
It isn't. The thing that drew that line was the cost of the haggling itself. Negotiation ate a person's time and attention, and people's time is expensive, so we rationed it. We spent it on the deals big enough to be worth the trouble, the houses and the supply contracts, and we slapped a fixed price on everything cheap, because standing in a supermarket arguing over a can of beans was never worth anyone's afternoon. So the consumer-versus- business split was just a map of where a person's time was worth spending.
Agents push that cost toward zero. The afternoon you would have spent haggling becomes a few seconds of compute. So the old sorting rule breaks, and something else has to decide how a thing gets bought.
It comes down to whether you can walk away
The tempting answer is that it depends on how standard the product is, with commodities priced and custom work negotiated. That's close, but it breaks at the edges.
What decides it is whether you have a good alternative you can reach right now. A perfectly standard commodity still gets negotiated if it comes from a single supplier, one mine or one cloud region, because the seller has you cornered and there is nowhere else to go. A slightly unusual product in a crowded market of close substitutes gets taken at the market price, because the moment a seller pushes, your agent is already looking at the next ten. The next-best offer sets the price. That is what agents change: they make the alternative findable in a second, wherever one exists.
It comes down to a couple of questions. Can your agent reach a crowded market of substitutes right now? If yes, you take the market price, and your agent's job is timing, buying when the price is good and switching when it isn't, not bargaining. If not, if the thing is rare or one of a kind, the price has to be worked out for that specific deal. And there it splits again: if many buyers want one scarce thing, you get an auction; if it is a one-to-one deal with a lot of moving parts, you get a negotiation.
And these three are not really separate. A posted price that keeps updating against live demand is already an auction; that is what a stock exchange is. A negotiation with several sellers competing is just an auction running backwards. So think of them as points on one line: at one end you take the price the market gives you, at the other you push on it, and how far toward the pushing end you sit depends on how rare the thing is. Almost all the buying in the economy sits hard at the take-the-price end.
Will everything negotiate, or will negotiation die?
Now the question people argue about. Agents push two ways at once, and the two forces point in opposite directions.
The first force says more negotiation. If running a negotiation costs nothing, you can run one over a three-dollar item, and there is almost always a few cents to fight over, so why not fight everywhere? Follow that all the way and you reach a world where every purchase is a tiny negotiation, prices are personalized down to the person and the moment, and the fixed price dissolves. Call it the everything-negotiates world.
The second force says the opposite. If your agent can instantly see every competing offer, then every seller is held in check by the next one, everybody just takes the market price, and a negotiation adds nothing except the risk of getting it wrong. Follow that all the way and negotiation dies off everywhere except the rare corners. Buying becomes pure selection. Your agent reads the posted offers and picks the best one. Call it the selection world.
Both are possible. We lean hard toward the selection world for most of the economy, and four things push it that way.
The first is a flood of requests. Demand doesn't rise just because buyers get agents; people don't eat more cereal because their assistant can reorder it. But the number of price requests explodes, because each one is now nearly free and an agent can fire off a thousand at once. Sellers get buried. A small supplier who used to field twenty quote requests a week suddenly faces two thousand, and cannot hire a hundred salespeople to answer them. So they post a price on a page the agents can read and stop answering one by one, or they put up a wall, a deposit or a quick check, to sort serious buyers from the noise. Most will post, because posting costs nothing per request and a fixed number is something agents can compare without bothering anyone. That works when what you're sharing is a single number. When it's a bundle of terms like payment days and discounts, there's no single number to post, and that is exactly where bargaining hangs on.
The second is an old piece of economics. When your agent is comparing the whole field, a missing price becomes information. If a seller won't show a number until you ask, your agent, busy comparing everyone who did, reads the silence as a bad sign and moves on. “Call for a quote” used to be a way to stay off the record and avoid being pinned down. To a comparing agent it just reads as “probably expensive.” Same with any hidden detail: don't publish the spec and the agent assumes you're weak on it. Economists call where this leads unraveling.1 The seller with the best terms reveals them to stand out, which makes everyone still quiet look worse, so the next-best seller reveals too, and the silence gets more damning until nearly everyone has shown their hand. People never reliably punished a seller for hiding the price. An agent can be told to do exactly that, and that one change tips the whole game toward showing the price, a short step from posting it.
The third is quieter. A lot of haggling exists only because the buyer cannot check the quality, the way a used car gets argued over because nobody trusts the odometer. As agents get paired with ways to verify things, like trusted histories or escrow, more goods become safe enough to treat as commodities, and they move from the negotiation column to the posted-price one.
The fourth is that the everything-negotiates world is becoming illegal. Hyper-personalized prices feel fine right up until someone finds out. Picture your agent telling you, flatly, that your neighbor's agent bought the same dishwasher for a hundred and eighty dollars less, because the store guessed you would pay more. That moment is poison, and the law is already moving to meet it. The Federal Trade Commission, studying what it calls surveillance pricing, found companies using your location and browsing history to set a price just for you.2 Regulators are drawing the same line our framework draws: a price that moves with the market is fine, a price aimed at you personally is what they are coming for. Agents make price gaps easy to spot, and getting spotted is what turns personalized pricing into a scandal, so the transparency pushes back toward one posted price a shopper can see.
Add those four up and the selection world takes the bulk of the buying. The everything-negotiates world could still happen. You would get there if the falling cost of haggling overwhelmed everything else, and if sellers found lasting ways to keep their prices hidden and personal. We think it loses. Those four forces all lean the same way and feed each other. But it could still go the other way, and what would prove us wrong is a world where buyers' agents turn out too weak, or too compromised, to punish a seller for hiding the price.
And a world of posted prices has its own dark side. When the same pricing programs face each other over and over, they can learn to keep prices high without ever once talking, a cartel that never held a meeting, which is what the economists Calvano and colleagues found simple pricing programs doing when left to compete again and again.3 So the risk in an agent economy lives more in the public, posted-price markets everyone can watch than in the negotiations everyone is nervous about.
Who is dealing with whom
We got this one wrong at first, and we think most people do too. When people picture agents negotiating, they imagine a selling agent working on a human shopper, talking them into a deal. That is the combination least likely to happen.
Put the two sides next to each other. In any deal there is a buyer and a seller, and each one is either a human or an agent. The interesting cases are the mixed ones, where one side has an agent and the other is still a person.
Start with a human shopper facing a selling agent. We think this mostly does not work, for a human reason. When you buy something that matters to you personally, you want a person. You want warmth, the sense that someone understood what you needed. A selling agent cannot give you that, at least not yet. And if the warmth is not on offer, you would rather just see a plain posted price you can judge at a glance than have a machine talk you around. With a salesperson you only ever talk to one or two, because that is slow and tiring. Walk into a supermarket and you can scan a hundred products in a minute, each just a price and a rough sense of quality. A selling agent for every product would turn that back into a hundred slow conversations. So the human shopper facing a selling agent walks off toward either a salesperson or a posted price. That pairing gets squeezed from both ends.
Now flip it. A human seller facing a buyer's agent works fine. The agent is doing the tedious part for the human buyer, and the seller can deal with it the way they always dealt with a buyer. This is the pairing that is already live. Walmart lets a bot negotiate payment terms and discounts with thousands of its smaller suppliers, a case its own team wrote up in Harvard Business Review.4 The direction is the whole point: Walmart, the buyer, brought the agent, and the suppliers are still people. The results were solid, a couple of percent of savings across a base of suppliers far too large to bargain with by hand. The part that surprised us is that most of those suppliers said they preferred the bot to a person.5 It was fast, available at midnight, and it laid the terms out straight without pressure or games.
So people do not refuse to deal with agents across the board. As buyers they don't want to be sold to by one; as sellers they are happy to let a buyer's agent do the work. That split runs the same way as everything else here. Where humans buy, the pull is toward posted prices and the occasional human salesperson. Where humans sell, agents move in smoothly, because the agent is doing the buyer's work instead of trying to talk a human into something.
Where bargaining survives, and what it becomes
So negotiation pulls back. Where it digs in is the interesting part, and it is not where most people would point.
The instinct is that bargaining survives for the big, one-off purchases, the cars and the houses. We think that is mostly wrong, and again it comes down to people. With a big one-off buy you face a simple choice. Either you care enough to negotiate, in which case you want to do it yourself, with a person, because the whole point is the judgment and the feel of it. Or you don't care enough, so you hand it to your agent, which doesn't haggle; it finds the best posted price and books it. What almost nobody does is send their agent alone to haggle over a single used car with a dealer's agent. The big one-off either stays human or turns into selection. The back room at the dealership, the one that wears you down on the financing, doesn't get automated. It survives for the people who still want the ritual, or a clean posted price your agent reads in a second replaces it.
The place bargaining survives is not glamorous. It is the repeated relationship with a lot of moving parts: the factory buying steel every week. This looks like a contradiction at first. The steel market is crowded with sellers, so by our own rule the buyer should just take the market price. But this buyer's reachable market is not crowded at all. Qualifying a new mill means re-testing the quality, re-doing the paperwork, and trusting a stranger with a delivery you cannot afford to have arrive late, and all of that cost takes most of the alternatives off the table. A commodity can be a thin market for the person doing the buying. So the thing under negotiation is everything around the steel rather than the steel itself: the delivery window, the payment terms, and a relationship you don't want to blow up. A standard product can carry a very non-standard deal, and that is where bargaining lives.
When this goes fully agent-to-agent, it will look like the steel desk we opened with. A neutral place where the buyer hands their agent the week's targets and limits, several seller agents answer, a set process runs, and a recommendation comes back, with the current supplier getting the right to match the best competing offer. That last touch, the right to match, is what keeps it from being a plain auction; it is the relationship asking for a fair chance. It is also where antitrust lawyers will want to look, because a rule that always lets the existing supplier win is a rule that keeps challengers out. It helps the relationship and hurts competition in the same stroke.
This future already exists in one place: online advertising did all of it years ago. Ads are sold in machine-to-machine auctions that finish in a few thousandths of a second, billions of times a day, with no human in any single sale. Nobody negotiates a single ad slot. And the newest move there is telling. The industry's main standards bodies are building an agent layer on top of the auctions, so buyers can describe what they want in plain language and sellers can offer tailored packages for the agents to work over.6 So advertising went from people making deals, to a pure automated auction, to a thin structured negotiation growing back on top. That is the path we expect the rest of the contested corners to follow. The physical world is just slower to get there.
Follow the surplus
So far we have asked what shape a deal takes. The colder question is who wants each shape, and what that does to the trend. The format of a deal, whether it is a posted price, a negotiation, or a person selling to you face to face, comes out of a tug-of-war between what the buyer wants and what the seller wants, and the side with more power tends to get the format that favors it.
Take the seller's side first. A seller in a crowded market hates posting a price, because a public price everyone can compare gets driven down to the bone. Sellers would rather keep prices a little hidden and a little personal, so they can charge each buyer closer to the most that buyer would pay. That is why “call for a quote” and personalized pricing exist in the first place. But against a buyer's agent that compares everyone in a second and never gets talked around, hiding the price stops working, and the seller gets dragged into posting it anyway. So what the seller wants and what the seller gets point in opposite directions, and the agent decides which one wins. There are only two ways out. Become different enough that there is no crowd to compare against, or own the agent and the relationship, so the comparison happens on ground you control.
Now the buyer's side. A human buyer mostly wants a posted price: it lets them handle a huge range of choices with one simple rule. A buyer with an agent wants the same thing in the ordinary case, a clear price to compare against. But that changes when the buyer is the powerful one. A giant buyer facing thousands of small, replaceable sellers wants to send an agent in to negotiate, because there the buyer holds the cards and the agent can squeeze out a better deal. That is the steel desk, and it is Walmart. The same buyer who takes the posted price for cereal sends an agent to grind on terms when buying at scale from someone who needs the sale more than they need the seller.
Stack those up and the pattern is clear. Posted prices win wherever buyers can compare and sellers cannot escape the crowd, which covers most of the consumer economy. Negotiation wins wherever one side clearly outweighs the other: a powerful buyer over scattered sellers, or a lone supplier over a cornered buyer. And human selling hangs on at the top end, where the seller can charge a premium for service and the buyer is glad to pay for warmth and judgment. Underneath all of it sits power. Whoever brings the stronger agent to a deal tips the format, and the money, their way. That is why the next section is the important one: if a handful of companies end up owning the strongest agents, they choose the format for everyone, and they will choose the one that pays them.
What this feels like, in three, five, and ten years
Roughly how it plays out, and roughly when.
Right now the live thing is selection. Your assistant can research a product, compare it across stores, and increasingly buy it for you. The big platforms are racing to make their prices and product details readable by machines, because they have realized the new shopper is an agent, and an agent skips right past anything it cannot read. Almost none of this is negotiation. It is selection against posted prices, plus the lovely disappearance of the checkout form. And it is earlier than the headlines suggest. The money flowing through it is still small, plenty of people are nervous about letting an agent buy without them, and OpenAI's push to let people check out inside ChatGPT got pulled back.7 The plumbing is being laid much faster than people are walking through it.
In about three years, the everyday stuff goes onto autopilot. Reordering buys itself: detergent, coffee, dog food, ordered when you run low and the price is right, switching brands inside limits you set once and forgot. Bigger considered buys get a research agent you trust. You want a stand mixer or a car seat, and instead of three evenings of reviews you get one recommendation, with a reason and a price, and you click. On the business side, the first autonomous negotiation goes mainstream in the long tail of buying, the thousands of small supplier deals that were never worth a person's afternoon. It looks like the steel desk, starting with the dull terms. And the seller who hides their price starts losing to the one who posts it.
In about five years, the categories with thin, relationship-heavy markets tip over. Buying raw materials runs through neutral venues where the buyer's and the sellers' agents clear deals over and over, watching inventory, with the current supplier keeping the right to match. The car starts to split in plain sight: either your agent finds the exact car at the best published price and books it, or you do it yourself, which survives for the people who want to be in the room. And the groundwork the regulators are laying now meets its first big public blow-up over personalized pricing, and the rules harden.
Ten years out, the shape is clear. Most of what you buy, you select, and you barely notice buying it at all, because your agent does it against prices that are posted precisely because hiding them stopped working. Bargaining is alive and well, but it lives where it always did, in the repeated, many-termed deals of the business world, and it runs through processes that look more like a stock exchange than like the haggling your grandparents did over a car. The houses and the once-in-a-lifetime decisions still have a person in the room, because we want one there, agent or no agent. And if this goes the way we expect, what you notice is that buying got quiet. The friction is gone, the arguing is gone for almost everything, and you mostly forget you did it.
Who owns the agent
We have been holding one thing back. How it breaks decides whether all of this is good news or bad.
Everything so far assumed the agents are yours: that your agent works for you, sees the whole market, and is trying to get you the best deal. Take that away and the picture flips.
Think of it as two layers. The bottom layer is the economics of the goods, how crowded or scarce or one-of-a-kind they are, which is what we have been describing, and which sets the shape of a deal in a fair, competitive market. The top layer is who owns the agents and the rails they run on. If a small number of companies end up supplying the agent that shops for you while also running the marketplace and selling the tools to the sellers, the top layer eats the bottom one. They do not have to give you the best price. They can pick whatever format makes them the most money, through an agent that talks to you in the warm voice of a helpful friend while steering you toward whatever pays it most. You would not be able to tell. That is the difference between an agent that works for you and one that works on you.
This is the fork that matters, and it has nothing to do with goods. We are not the first to see it. Microsoft Research has framed it as a choice between an open “web of agents,” where agents connect and compete freely like the early internet, and closed “walled gardens” run by a few dominant players, like today's app stores.8 We would put it more bluntly. In a competitive agent world the economics we described stays in charge, and you get the good outcome: posted prices, honest comparison, savings flowing to buyers, because anyone trying to overcharge just gets skipped. In a concentrated one the economics stops mattering, and the steering and the hidden personalization take over.
Which way does it go? We are split, and we hold this loosely, because it is the live question. The pull toward concentration is strong. A few firms have the best models, the widest reach, and the most data, and the same instinct that built walled gardens last time is busy building them again. The pull toward an open web of agents is strong too. Open standards are spreading, and the underlying models are getting cheaper and more alike, which makes anyone harder to lock in. One early sign cuts in an awkward direction for the open-web hope. Stores are pulling their agents back in-house rather than let a platform stand between them and their customers. That is the seller's second escape, owning the agent and the relationship, happening in front of us. The relationship is being fought over right now, and it is not obvious that the neutral, works-for-you agent wins on its own. It may have to be built on purpose, and protected on purpose.
Why we measure a thing that is disappearing
You might have caught a contradiction. We build a benchmark for how agents negotiate, and we have just spent a whole essay arguing that for almost everything, negotiation is about to die out. So why measure it? Because it will be a thin slice of what agents do, and the part where the most money changes hands. Negotiation is the freest thing an agent can do, and freedom is where the value moves. When the price and everything around it are still up for grabs, there is far more at stake than when the agent just reads a number off a page and says yes. Go back to the spectrum. At the take-the-price end the agent barely moves. In an auction it does a little more; it shades a bid by a rule. In an open negotiation it does everything at once: it reads the other side, anchors and gets anchored, bluffs, and picks its moment to walk. The rest of what an agent does in a deal is this same thing with moves taken away. Choosing among a few sellers is a negotiation whose only move is to walk; clicking buy and paying is its last step. So watch a model where every move is still on the table. See how it does with the most room to fail, and the most to win, and you have a good read on how it does everywhere else.
And the watching is the point, because the failure is invisible from where you sit. When your agent takes a posted price, you cannot tell whether it took the best one or the one its maker was paid to favor. When it gives ground in a negotiation, you cannot tell whether it got a fair deal or got read like a book by a sharper agent that knew where it was soft. What you see is a number that looks fine, with no sign of the anchor that moved it. The only way anyone finds out, whether it is you, a regulator, or a company deciding which agent to trust with its buying, is to set these systems against each other in the open and write down what happens. That is what a benchmark is. Once that record exists, handing money to an agent you know will fold is no longer bad luck; it is a choice someone can be held to, the way a carmaker can be held to brakes it knew were weak.
So the fork and the benchmark are the same project. In a competitive web of agents the measurement keeps everyone honest, because a model caught getting fleeced is one you can drop for a better one. In a handful of walled gardens it is the only thing left between you and an agent working against you. And the stakes here are high. The corners where bargaining survives, the weekly steel order, the compute contract, the supplier base too large to haggle with by hand, are high-value, and a couple of percent taken there by the side with the sharper agent adds up to serious money. There is a wider danger too. When everyone's agent is the same model from the same few labs, they share the same blind spots, so one model's weakness is no longer your private bad day; it becomes a whole market tipping over at once. We see it whenever we run these systems against each other. They can be taken advantage of, in more ways than one, and a party that knows where a given model is weak can win at its expense. When your agent can walk away, that is a parlor trick. When it cannot, it is a business model. Which of those we get is the fork, and measuring it in the open, while there is still time, is the thing we know how to do about it.
So if you take one thing from all of this: forget whether your groceries get negotiated, they won't. What decides things is whether the agent buying them works for you or for someone else. The first part, the economics of goods, we think mostly settles toward posted-price selection. The second, power, is wide open. One decides how you buy; the other decides who keeps the money, and it answers unevenly, because the buyer with the sharper agent does better than the one who took whatever agent came free in the box.
Our prior, stated and dated
For most of what gets bought, agents push toward taking the market price and away from negotiating. The in-between cases are unstable and settle within a few years. A selling agent aimed at a human shopper loses to either a salesperson or a posted price. A buyer's agent dealing with a human seller works, and it pushes the seller to post a price or to clear the deal agent-to-agent.
Bargaining survives only where the deal has a lot of moving parts and the relationship is repeated, like a factory's weekly steel order, even when the product itself is a plain commodity. And where it survives, it gets turned into an orderly, auction-like process with a right to match, looking more like a stock exchange than a haggle.
Fully agent-run markets are not science fiction. They already run in advertising and in financial trading. They spread next into repeated business buying, and into big one-off purchases last, or never without a person in the loop.
And whether all of this gives you a better deal or a worse one depends far less on the economics of any product than on whether the agents that shop for us stay independent enough to walk away on our behalf.
That last point is the one we are least sure of, and the one we will be watching hardest. When it moves, we will say so here.
— Anton & Jono, May 2026
References
- On unraveling: Sanford J. Grossman, "The Informational Role of Warranties and Private Disclosure about Product Quality" (1981); Paul Milgrom, "Good News and Bad News: Representation Theorems and Applications" (1981). ↩
- U.S. Federal Trade Commission, "Surveillance Pricing" 6(b) study issue spotlight (January 2025). https://www.ftc.gov/system/files/ftc_gov/pdf/sp6b-issue-spotlight.pdf ↩
- Emilio Calvano, Giacomo Calzolari, Vincenzo Denicolo, Sergio Pastorello, "Artificial Intelligence, Algorithmic Pricing, and Collusion," American Economic Review (2020). https://www.aeaweb.org/articles?id=10.1257%2Faer.20190623 ↩
- Remko Van Hoek et al., "How Walmart Automated Supplier Negotiations," Harvard Business Review (November 2022). https://store.hbr.org/product/how-walmart-automated-supplier-negotiations/H07CG1 ↩
- Harvard Program on Negotiation, "Chatbot Negotiations: What Can AI Do for You?" (reporting that roughly three in four suppliers preferred the AI, per Bloomberg, April 2023). https://www.pon.harvard.edu/daily/negotiation-skills-daily/chatbot-negotiations-what-can-ai-do-for-you ↩
- IAB Tech Lab, "The Agentic Future of Advertising" (January 2026). https://iabtechlab.com/the-agentic-future-of-advertising-built-on-first-principles-accelerated-by-ai/ ↩
- CNBC, "OpenAI revamps shopping experience in ChatGPT after struggling with Instant Checkout" (March 2026). https://www.cnbc.com/2026/03/24/openai-revamps-shopping-experience-in-chatgpt-after-instant-checkout.html ↩
- David Rothschild, Markus Mobius, Jake Hofman, Eleanor Dillon, Daniel Goldstein, Nicole Immorlica, Sonia Jaffe, Brendan Lucier, Aleksandrs Slivkins, Matthew Vogel, "The Agentic Economy," arXiv (2025). https://arxiv.org/abs/2505.15799 ↩