Why My Agent Punched the Monkey

This week, I have been thinking about the monkey because Amazon just threw Meta’s new Muse agent out of its store and blocked the bot from perusing the aisles or even crossing the virtual threshold. You remember the monkey, right? It lived in a banner ad around the turn of the millennium and his job was to sit there looking punchable until you punched him. Punch the monkey and win! Perhaps an iPod. Perhaps $20. Perhaps absolutely nothing except the privilege of discovering that you had just volunteered your email address to seventeen mortgage brokers in Florida. It was stupid, magnificently stupid, but the monkey understood something fundamental about the internet before most of us did: the product was never the monkey. The product was your attention. We have spent the quarter-century since then constructing the most sophisticated machinery in human history for getting that attention, measuring it, auctioning it, following it around and eventually selling it all of the things.

This attention management land grab turned out to be an almost preposterously good business. Google discovered that instead of dangling monkeys, you could wait until somebody typed “cheap flight to Paris” and sell Delta the privilege of standing closest to the keyboard. Later, Facebook took it up a notch and discovered that you didn't even need to wait for people to want something. Show them a sufficiently beautiful person wearing sufficiently beautiful shoes in Tuscany and desire could be manufactured wholesale, teaching us all to desire the things we never imagined we needed. More on that, and Meta Muse, later.

The result? A whole lot of bananas at a very high margin . In 2025, Google advertising generated $294.7 billion, almost three quarters of Alphabet's revenue, while Alphabet produced $164.7 billion in operating cash flow. Meta took the proposition even further: advertising produced $196.2 billion of its $201 billion in revenue, which is about as close as a $200 billion company gets to saying we sell ads. Amazon pulled in another $68.6 billion from advertising, up 22% in a year, while Walmart's much younger global ad business hit nearly $6.4 billion, up 46%. Advertising helped turn a search box and a social network into two of the great cash-generating machines in corporate history, and now it is throwing off billions for companies that supposedly sell groceries and cardboard boxes. Nor is this some elderly industry wheezing toward its dotage. U.S. internet advertising hit $294.6 billion in 2025, up 13.9% in a single year, with commerce media alone reaching $63.4 billion. 

There is a funny thing that happens to companies once they discover advertising. They start doing more of it, and eventually the thing they originally built begins quietly rearranging itself around the ad business. This is understandable because selling advertisements is an almost indecently attractive activity compared with selling actual things. Actual things must be manufactured, shipped, refrigerated, insured, returned, occasionally recalled and sometimes repackaged. An advertisement does not require a refrigerated truck, and nobody has ever returned one because they changed their mind.

Walmart is a wonderful example because Walmart is, ostensibly, a company that buys physical objects, moves them through warehouses, puts them in stores and sells them for slightly more than it paid. Then Walmart discovered that Procter & Gamble would also pay Walmart for the privilege of putting Tide in front of the person who typed laundry detergent. To be fair, grocers have long charged lucrative stocking and placement fees for physical stores. But just as Amazon introduced the endless aisle, so too did Walmart discover the endless advertising inventory to match. To wit, the world’s largest retailer’s global advertising business reached nearly $6.4 billion in fiscal 2026, up 46%. 

Apple caught the same ad mania, although in the characteristically Apple fashion of declining to tell us exactly how much money the disease is making. Apple reports advertising inside Services rather than breaking it out separately, but says advertising helped drive Services growth in fiscal 2025. Services generated $109.2 billion in revenue and $82.3 billion in gross margin, which works out to a gross margin above 75%. Apple designs some of the most complicated consumer hardware on Earth, involving custom silicon, glass, aluminum, factories, ships, stores and enough supply-chain choreography to invade a medium-sized country. Then somebody pays to put an app in front of you in the App Store and, financially speaking, everyone gets to go home early. Anecdotally, this is a high margin and high revenue business.

Amazon, naturally, has made a science of the whole thing. Its original retail miracle was logistics: enormous warehouses, planes, vans, robots and human beings sprinting around fulfillment centers so a USB cable can materialize on your porch before you've completely accepted that you ordered it. And they managed to get paid well before they had to pay their suppliers, pocketing the float.  

But sitting on top of this industrial marvel is something much cleaner. Amazon sold $68.6 billion of advertising in 2025, up from $56.2 billion the year before and $46.9 billion the year before that. Amazon doesn't disclose the operating margin of advertising separately, so anyone giving you a precise figure has wandered from accounting into fan fiction, but the company does say increased advertising sales contributed to higher operating income. 

In a nutshell, it’s about this.  When you own the aisle, the shelf, the search box and the cash register, sooner or later somebody realizes you can rent them out and that’s a better business than selling the thing.

Now give me an agent programmed to do my bidding (or so we hope) and tell it, "Get me another pair of the black size 10½ shoes I bought last year, delivered by Friday, and don't pay more than $160." There is no particular reason for the agent to look at a search-results page, sponsored products, a carousel or customers who bought this also bought an air fryer. “Amazon’s choice” becomes meaningless. And the agent doesn't need to wonder why the first six results for the exact product I requested somehow aren't the exact product I requested. The whole commercial pageant can collapse into one computer buying shoes from another computer. The agent is an ad blocker with a checking account, and unlike the browser variety, this one can finish the transaction without ever showing me the page it blocked.

That is the real reason, IMHO, why Amazon blocking agents is a much bigger question. There are the standard arguments about bots, scraping and terms of service. Some of those complaints are perfectly legitimate. Nobody expects Amazon to let badly behaved software wander through its systems using stolen credentials or pounding servers without limits. But that’s not what Perplexity and Muse were doing. They were agents acting on behalf of legitimate users who gave them explicit permission to go shopping on Amazon. And so, there is an elephant standing behind all of this wearing an Amazon Basics T-shirt. Amazon has built a $68.6 billion advertising business because it owns the final few feet between I want something and I bought it, while Walmart is racing down precisely the same road because those final few feet are fantastically valuable. An agent doesn't have to destroy either company's retail business to cause trouble. It merely has to make those feet shorter and disrupt the flow of endless advertising opportunities that fueled the magical online money machines.

Almost $300 billion of American digital advertising isn't simply going to vanish. The obvious answer is that advertisers will pay the agents a commission or something else. Suppose I say, "Buy me the best noise-canceling headphones under $400," and Sony offers my agent $12 if it selects Sony while Bose offers $15 and Apple offers $20. Congratulations: we have invented Google Search again, except this time the sponsored result can operate my credit card. Note the tension. Who is the agent working for in this case? And if the company operating the agent is also on the other side of the transaction, selling the ads or the commissions or whatever, who is the agent really working for?

I mean, it is supposed to be my agent. But running one isn't free, particularly once it remembers who I am and starts doing things for me all day. Somebody eventually gets an invoice, and the internet's traditional answer to the question of who pays has been wonderfully simple: not you. Google is free because advertisers pay Google. Instagram is free because advertisers pay Meta. Amazon Prime isn't free, but Amazon has nevertheless discovered that customers who already pay admission can also be advertised to once they're inside, which is a little like charging someone admission to a casino and then renting the roulette wheel to Pfizer. The problem becomes stranger with an agent because an Instagram ad is trying to manipulate me, while an ad inside my agent is trying to manipulate the thing I hired to do my manipulating for me. Somewhere Robert Sapolsky is smiling. We have finally outsourced the dopamine loop and made it an infinite mess.

There is another possibility. Maybe advertisers don't pay agents to recommend things; maybe they pay them to bring offers. This sounds like a distinction invented by a pYC pitch-deck team of B-school drop outs after three martinis, but bear with me. Tell my agent I want Jordans and Nike cannot persuade it that I actually need loafers. Foot Locker cannot purchase a more emotionally resonant shade of blue because there may be no screen at all. But merchants can compete: $145 here, $151 with same-day delivery there, $155 with 5,000 loyalty points or $160 with free returns for 90 days. What used to be an advertisement designed to make me feel something has turned into somebody offering me a better deal — a reverse auction, if you will.

That sounds wonderful for consumers until you notice what we've done to the merchants. We have constructed an immortal, infinitely patient procurement officer who can ask every shoe seller on Earth for its best price at the same time, never gets embarrassed about asking for another discount and will cheerfully abandon a retailer over $3.72. Retailers will adore this in approximately the same way taxi drivers adored Uber. If a large percentage of commerce starts this way, a great deal of what we currently call performance advertising begins to look unnecessary. Why bid $4.73 for my click so you can offer me a shoe for $160 when my agent can simply ask whether you'll take $148?

This is where the monkey gets his revenge, because if you can't advertise to me after I know what I want, you have to get to me before I know what I want. Consider the difference between “show me stuff from the latest Paris Fashion Week” and “buy me Jordans.” There is an entire advertising industry hiding between those sentences. The first request gives an agent enormous latitude: Zegna originals, Zara knockoffs, some obscure Japanese designer I have never heard of, or whatever else it thinks the version of me represented in vector space might find irresistible. The second is procurement. Nike has already gotten the word Jordan into my head, so the agent isn't choosing the brand anymore. It can set every retailer selling my size against every other retailer selling my size, and the Borg is on my team.

This is the part of the story where Meta starts looking less like another advertising company standing in front of the agent steamroller and more like the company that may already own a large piece of the road. Instagram doesn't merely harvest demand; it creates it and has spent years trying to follow it toward the cash register. Google waits for best hotel in Kyoto. Amazon waits for Sony WH-1000XM7. Instagram shows you a photograph of a ryokan or a pair of headphones you had no intention of buying until fifteen seconds ago. Nobody goes onto Instagram and types purchase aspirational linen shirt. You see an implausibly attractive Italian — probably AI-generated by now — drinking an espresso beside Lake Como and discover, several neurons later, that apparently you have always been a linen-shirt person. Meta generated $196.2 billion of advertising revenue in 2025, up 22%, and the company says online commerce was the largest contributor to that increase. Whatever agents do to advertising, they don't repeal the human capacity to see a handsome stranger wearing a shirt and suddenly need the shirt.

That is what makes Muse more interesting than a bot that waits obediently for a shopping list. Meta already owns much of monkey see, monkey want, monkey do. Instagram can show me the jacket I didn't know existed, convince me that my life has somehow been impoverished by its absence and send me toward the cash register. Muse potentially takes out the seams. What's that jacket? becomes would it look good on me? becomes find it in my size becomes don't pay more than $300 becomes a package on my porch. Google and Amazon have pieces of that chain, but Meta already owns one of the world's great machines for turning people who weren't shopping into people who suddenly are.

And Meta has been training for this particular fight for years, partly because Apple gave it no choice. When Apple's privacy changes blew up mobile tracking, Meta said the hit would cost it about $10 billion in 2022 alone. It responded by getting much better at recommending things with less knowledge of what users had done elsewhere. That turns out to be a useful skill if the next great advertising surface isn't a webpage at all, but an AI deciding what you might want before you have asked for it.

There is an uncomfortable question buried here about whose side Muse is on during the handoff. If Meta was paid to put the jacket in front of me and Muse then offers to buy the jacket for me, at what precise millisecond did it stop working for the advertiser and start working for me? Perhaps it never does, and perhaps most of us decide we don't particularly care if the deal is good enough. Human beings have tolerated considerably more compromised arrangements in exchange for free email, search and video. Meta doesn't necessarily have to defend today's advertising funnel if Muse works. It can put discovery and purchase into the same product and eat the funnel instead.

Here is where this stops being a story about the future of advertising and starts becoming a story about the future of some of the most profitable companies on Earth. Agents don't have to kill advertising for the profit math to get ugly. Suppose they merely remove some of the sponsored searches and promoted products that appear after I've already decided what I want. The lost dollar isn't necessarily an average dollar of revenue. It can be some of the nicest money these companies make. Amazon doesn't disclose advertising margins separately, which prevents us from putting a clean number on this, but it explicitly tells investors that advertising contributes to operating-income growth. Walmart similarly points to advertising as improving its business mix. Take enough of that cream off the top and you can change the economics of a company without making much of a dent in its headline revenue.

That helps explain why Amazon isn't greeting autonomous shopping agents at the front door with cookies. Nobody knows how much ad spending agents can actually displace, and anyone handing you a tidy forecast for 2030 should probably be made to wear a sandwich board saying I have seen the future. But you don't need a forecast to see the problem. If software starts taking possession of the trip from I want that to I bought it, then owning the webpage along the route becomes less valuable. For a generation of companies that became fantastically profitable by charging rent on that trip, even losing a few exits could hurt. 

To be crystal clear, Google, Walmart and others are running in the other direction, embracing agentic commerce. This makes sense, considering that Amazon remains the biggest player in online commerce and the one with the most at stake. It’s entirely possible that the companies running pell mell to bring agents into their stores are following the same complement as disruption story that Google executed to a tee with its Kubernetes container orchestration gambit to combat Amazon’s early cloud computing dominance. 

This is where Ben Thompson's Aggregation Theory gets entertaining, because agents aren't merely another interface sitting on top of the web. The plumbing underneath them is changing too. MCP has become a standard way for agents to connect to tools and services, while Google's Universal Commerce Protocol is aimed directly at the shopping problem: giving agents and merchants a common language for discovery, checkout and order management. UCP can itself work over MCP, A2A or ordinary APIs and was developed with companies including Shopify, Etsy, Wayfair, Target ajnd Walmart, with more than 20 other companies endorsing it. Google has also built AP2 for agent payments, and there is already enough protocol alphabet soup forming around agents to make TCP/IP look refreshingly literary.  Amazon itself has embraced MCP and some of these other protocols because it recognizes that, in the future, agents will conduct transactions and interact autonomously. 

In the dawn of the new agentic daylight, the future doesn't have to be an increasingly sophisticated bot wearing a Chrome moustache and pretending to be a human so it can scrape Amazon. A merchant can expose what it sells, what it charges, whether it can deliver Friday and how to buy it informs what agent can understand directly. UCP explicitly covers the trip from product discovery through checkout and order management, including discounts, and lets agents dynamically discover what a merchant  supports. Once enough merchants do that, the agent doesn't necessarily need Amazon's homepage. It needs merchandise. If Amazon won't supply it on terms the agent likes, the agent can ask someone who will.

This is where Curious George gets into the control room. Google aggregated websites, Meta aggregated attention and Amazon aggregated merchants because they controlled the route through which consumers reached enormous numbers of suppliers. My agent can know what shoes fit me and which airlines I hate, remember that I'll pay $80 more for a nonstop flight but not $200, and then go wandering through MCP, UCP, A2A and whatever protocol required to find someone willing to sell me what I want. From my agent’s perspective, Amazon and Walmart are places to get merchandise, Expedia is somewhere to get fares and a hotel chain can answer for itself. For twenty years, the great internet companies fought over who got to stand between you and the thing you wanted. Protocols give Curious George an endless aisle in a single menu, zero friction and the tools to disintermediate everyone who made a living standing between the buyer and their purchase.

So where does the money go? At the bottom of the funnel, today's sponsored listings could turn into a vast machine-to-machine bazaar in which my agent announces that I want the Jordans and retailers to respond with prices, delivery promises, return policies, loyalty points and perhaps a commission for whichever agent brought the customer. The commission is where things get slippery. If my agent pockets $8 from Foot Locker but Foot Locker still gives me the best deal, perhaps nobody cares. If that $8 causes it to quietly ignore a better offer from Dick's, we have rebuilt the sponsored search result inside a machine that claims to work for me. The ad hasn't disappeared. I just can't see it anymore.

That may be the basic bargain behind the free agents most of us eventually use. Advertisers and merchants subsidize them, just as they subsidized search, social media and much of the consumer internet before them. But instead of buying pixels, they buy some kind of access to the agent: the right to submit an offer, a commission when a transaction closes, perhaps a fee to be considered at all. The clean version looks a lot like affiliate marketing with a terrifyingly competent negotiator in the middle. The dirty version looks like Amazon search without the little Sponsored label. Somewhere between the two is probably a business model worth an obscene amount of money.

Farther upstream, the money goes hunting for desire. Nike cares less about winning the agent's shoe auction if Nike can make sure I ask for Jordans in the first place. Coca-Cola doesn't need to persuade my grocery agent if I told it to buy Coke. Luxury goods become particularly funny because the entire point of a Birkin bag is that no rational procurement algorithm would ever invent one. A lot of the money now spent catching us on the way to a purchase could move toward Instagram, TikTok, creators, sports, entertainment and all the other places where humans acquire wants they cannot explain. And toward Muse and other agents, where we talk to our Ids and describe the person we are absolutely going to become once we lose fifteen pounds, learn Italian and finally buy the right jacket. Sapolsky gets the monkey. The agent gets procurement.

There will also be agents that refuse the whole arrangement. Pay $30 a month and your little electronic Jeeves takes no commissions, accepts no sponsored offers and promises to regard every merchant with the same chilly indifference. I would probably pay for that. Most people probably won't, because we have already run this experiment with search, email, news, social networks and video, and free has beaten “better aligned with my interests” often enough that I would place a Kalshi contract on the outcome. The mass-market agent is therefore likely to have somebody else somewhere in the transaction paying the bill.

Which leaves us with a peculiar possibility. Advertising doesn't die at all. It becomes harder to see. The sponsored tile turns into an offer passed to an agent; the affiliate link becomes a machine-readable commission; the coupon becomes a bidding mechanism; the ad auction becomes a negotiation among software that takes place before I know there was anything to negotiate. We will undoubtedly invent a magnificent name for this — agentic commerce optimization or autonomous demand orchestration — and spend several years explaining that it isn't advertising.

OK, but please show me the monkey. If somebody is paying my agent to put something in front of me, I want to know who is holding the banana. Because, at the end of the day, I can slip on the peel if I don’t see it on the ground. Then my agent will probably have to punch the monkey.