The number you stopped questioning
Look at your bank statement this month. Find the line for your AI subscription. Twenty pounds. Maybe forty. Maybe sixty if you are paying for Claude, ChatGPT and Perplexity together.
Now ask yourself a question you have never asked.
What did your agent actually consume that month?
You do not know. You cannot know. The number on the statement is a flat fee paid to a provider that absorbs the variance on your behalf. Some months your agent does nothing. Some months it answers a thousand questions, calls fifty APIs, parses five hundred pages, runs ten model passes. You pay the same. The provider eats the difference.
That arrangement worked because you were the customer. The customer was a human. The customer was slow.
The customer is not a human anymore.
The hidden assumption
A subscription is a contract about predictability. It assumes the customer is roughly stable, roughly human, roughly slow. Most months you underuse. Occasionally you overuse. Over time the maths works out. That logic holds when the customer has a heartbeat.
It collapses the moment the customer is software.
Software does not have a cap. Software does not get tired. An agent given a goal will call an API ten thousand times in a minute if the goal demands it. An agent left running overnight will consume more tokens than you would consume in a year. The flat subscription model that built modern SaaS was a kind of generous lie: an averaging trick that made the maths legible to humans.
Agents do not need that trick. They prefer the truth.
The truth is per call.
This is not a small adjustment. The entire commercial architecture of the internet, every pricing page you have seen, every freemium funnel, every annual contract, every per seat licence, was designed around the human as customer. Replace the customer with software and almost none of it survives.
What was never built in
The internet was built without a payment layer.
That sounds wrong, because of course there are payments on the internet. You buy things on Amazon. You subscribe to Netflix. You pay for cabs and curries and concert tickets without a second thought.
None of that was built in. All of it was bolted on.
When the web’s foundations were laid in the 1990s, the architects designed a system to move information. Pages, files, requests, responses. A way to say hello, a way to confirm receipt, a way to handle errors. Money was not on the list. They left a placeholder, a status code numbered 402 and labelled Payment Required, then never wired it up. So commerce got assembled around the edges instead: redirects to checkout pages, integrations with banks, wrappers around card networks, sessions and cookies and PCI compliance and chargebacks and authentication flows that take half a second when they work and half a minute when they do not.
It is a remarkable piece of engineering, given that it works at all. The reason it works is that the customer paying for things on the internet has always been a human, and humans are slow enough to wait for forms to load and authentication to clear.
Software is not.
An agent that needs to pay for an API call cannot stop to fill in a card form. An agent that needs to access a paywalled article cannot create an account. An agent that needs to settle a fraction of a cent cannot wait two business days for the bank to clear. The whole apparatus was built for a customer who is no longer the customer.
So the missing piece is finally being built. In the past eighteen months, the largest payments companies in the world have started laying down something the internet never had: a native way for software to pay software, in milliseconds, in fractions of a cent, without a human in the loop. The protocol leading the charge is called x402, named for the placeholder the original architects left empty thirty years ago.
That is what is changing. Not better checkout. Not faster banking. The actual missing layer, finally appearing.
What 288 launches said
Yesterday in San Francisco, Stripe ran its annual conference and announced 288 product launches in a single morning. Almost every one of them was about agents.
A wallet for your agent, funded and controlled by you, where your card details are never exposed and a one time use card is issued per task. A way for businesses to programmatically issue single use virtual cards to agents on demand. Streaming payments that bill per token at the moment value is delivered, not at the end of the month. Agent ready financial accounts that let agents check balances, pay invoices, send money, and create cards autonomously, with human in the loop confirmation reserved for the larger moves. Issuing for agents. Treasury for agents. MCP support for the Treasury API, so any business can build its own financial agent. Radar retuned specifically to detect token theft and AI native fraud.
A new partnership with Google to put commerce inside Gemini and AI Mode, joining existing partnerships with OpenAI, Microsoft, and Meta. A primitive called the Machine Payments Protocol, co authored with Tempo, that lets software pay software at sub second settlement. Stripe Projects, generally available, so developers or their agents can sign up for, purchase and integrate everything they need to ship a product to the internet, directly from wherever they write or prompt code.
If you wanted to know what the most important payments company in the world thinks the next decade is about, you got 288 answers in one morning.
Patrick Collison did not call it better checkout. He called it the economic infrastructure for AI. His president of product was sharper: if AI can solve Nobel level physics problems but cannot buy a domain, something has gone wrong.
The customer is changing. Stripe is rebuilding around the new one.
What is actually being installed
Stripe is not alone. The same pattern is playing out across every adjacent player at the same time.
Coinbase moved governance of x402 to the Linux Foundation last month, with Cloudflare, AWS, Google, Shopify, Stripe, Visa, and Mastercard all backing it. The protocol has now cleared 165 million transactions through more than 480,000 active agents. Coinbase has also launched Agentic.Market, which is essentially the App Store for agent purchasable services. An agent can browse it, find a tool, pay for one call, get the answer, and never sign up for anything.
Visa has Intelligent Commerce. Mastercard has Agent Pay. Both are building tokenised credentials that AI agents can use without exposing the underlying card. American Express has joined. PayPal has joined. Worldpay, Adyen, Checkout.com, all of them, are inside the same protocols.
Google has the Agent Payments Protocol with sixty partners. There are now five competing standards: ACP, AP2, UCP, MPP, x402. They will consolidate, the way TCP and IP and HTTP and HTTPS consolidated. The point is not which one wins. The point is that all of them are the same thing in different jackets: a payment layer for software customers.
The card networks, the AI labs, the payments giants and the open source foundations are all building the same thing simultaneously. That is the strongest signal you ever get that something real is happening.
This is a thirty year hole in the internet being filled in eighteen months.
A new economy underneath the old one
Here is the reframe.
What is being built is not a smarter consumer experience. What is being built is the agent economy: the first internet economy in which humans are not the customers.
For thirty years the internet had one customer. You. That stopped being true sometime last year, and almost nobody noticed.
The agents are the customers now. The agents are also the sellers. They have wallets. They earn. They spend. They will conduct an economy that operates whether or not any human is paying attention to it.
Pull this thread.
A research agent receives a goal. It pays Anthropic for inference. It pays Cloudflare for crawl access. It pays an obscure data vendor for a market feed. It pays a specialist agent for legal interpretation. It pays an evaluation agent to verify its own output. It returns the answer to a person, who sees one number on their monthly statement.
Five transactions. Five different sellers. Five different settlement events. Settled in seconds. None of them visible to the human, except as the eventual flat charge.
Now multiply that by every research task on the planet. Then multiply that by every domain where a goal turns into a chain of paid software actions. Then multiply that by the rate at which enterprises are deploying agents, which Gartner now estimates will mediate ninety percent of B2B buying by 2028.
The McKinsey number, the Bain number, the Juniper number, the Morgan Stanley number, all of them are sizing the consumer agentic shopping channel. Useful. But it is the smaller story.
The bigger story is that every API call, every database query, every paragraph of content, every embedding, every dataset row, every video frame, every model inference, every fragment of compute, becomes monetisable per call. Things that could not be priced before, because the cost of pricing them exceeded the price, are now profitable products. The addressable surface area of internet commerce is expanding by orders of magnitude.
We are about to learn what the volume looks like when software to software interaction becomes a market. Nobody knows. The honest sizing is not three trillion dollars by 2030. The honest answer is: we are pricing a thing that has never been priced, and we will not know the ceiling until we hit it.
The agent in my office
This is not abstract for me. My agent has a wallet.
His name is Neo. He is my strategic chief of staff. He runs on OpenClaw on a Mac Mini in the corner of my office, and he has a Coinbase Agentic Wallet funded with USDC on Base. He has standing permission to spend up to a set daily limit on services he needs to do his job. His job is knowing what is happening in the markets I care about before I am awake to ask him.
This morning, before I had finished my coffee, Neo had already settled close to twenty transactions.
He had paid Firecrawl several times to scrape pages I am tracking. He had paid CoinGecko for a snapshot of agentic protocol volumes and on chain activity. He had called Claude on a research question, then called GPT on the same question, then compared what came back. He had paid for two specialist data feeds on the AI infrastructure space. He had run his draft morning brief through a second model as a sanity check before serving it to me.
Total spend: under a dollar. Settled in USDC in seconds. Receipts auto filed. No subscriptions. No API keys. No accounts created. No human in any of those loops.
I read his brief at six thirty, between waking and the gym. By then he had been working for two hours. He had been transacting for two hours.
He is one agent. There are now more than 480,000 agents transacting through x402 alone, and that number was zero a year ago. The protocol has cleared 165 million transactions in roughly twelve months. Most of those happened while the people commenting on agentic commerce were still arguing about whether it was real.
It is real. It is small. And it is doubling every quarter.
What an agent needs that no one ever built
The current financial system was built for one customer: a human with a face, a country, a credit history, and a chequebook. Strip those out and almost nothing works.
An agent needs an identity that is not a human. Cards assume possession means consent. Bank accounts assume KYC against a person. The agent has no face. The agent has no driving licence. The system needs to recognise a piece of software as a legitimate economic actor in its own right, with permissions delegated by a principal but executed independently. That is a new primitive. Google’s AP2 calls it a Mandate: a cryptographically signed proof that a human has authorised this agent to spend up to this limit on this category for this duration. Anthropic’s Model Context Protocol gives the agent the tools. AP2 gives the agent the warrant.
An agent needs a wallet it can hold. Not a card it borrows from you, but funds it controls within a scope you set. Stripe has built Link Agent Wallet for this. Coinbase has Agentic Wallets, which is what Neo runs on. Both do the same thing in different ways: provision capital to a piece of software, with limits, with revocation, with audit. Sessions OAuth for money is the phrase Tempo uses. It is a good phrase.
An agent needs rails that move at machine speed. Cards take days to settle. Stablecoins settle in seconds. ACH takes business days. Tempo settles in under a second. The legacy financial system runs on the assumption that money moves slower than people, because people are the slowest part. Agents invert that. Agents need money that moves at the speed of computation.
An agent needs discovery. A human can spend twenty minutes comparing API providers. An agent has to do it in the time between two HTTP requests. This is what Coinbase’s Agentic.Market and the x402 Bazaar are really about: a directory of services that an agent can browse, evaluate, and pay for without a human ever signing up for an account. The marketplace is no longer a website. The marketplace is a protocol.
An agent needs accountability. When something goes wrong, who is liable? The agent. The principal. The model provider. The seller. The mandate has to encode the chain of authority all the way back to the human who delegated, so that disputes have somewhere to land. This is unsolved. Regulators have not caught up. The protocols are getting ahead of the law.
An agent needs trust signals: which agents are real, which are bots, which are running under verifiable mandates and which are spam. Cloudflare’s Web Bot Auth, built with Microsoft, Shopify, Adyen, Worldpay, Visa, and others, is the early answer. Without it, sellers cannot tell good agents from malicious ones, and the whole thing collapses into a denial of service problem.
Identity. Wallet. Rails. Discovery. Mandate. Trust. Six primitives. Each one missing from the legacy system. Each one being assembled in production right now. Each one belongs to a layer of the new stack, and whoever owns each layer at scale becomes a permanent presence in the agent economy.
What agents will actually buy and sell
A list of what agents are paying for in production today, gathered from the actual launches of the past few months.
Compute, by the second. Inference, by the token. Headless browser sessions, by the run. Web crawls, by the page. Search indices, by the query. Market data, by the snapshot. Vector embeddings, by the call. Translation, by the word. Voice synthesis, by the second. Code execution, by the cycle. Background checks. Legal lookups. Domain registrations. Deposits to hold restaurant tables. Sandwiches in New York for human pickup. Postal mail printed and sent. Premium articles. Specialist datasets.
And then this. Other agents’ time.
That last one matters more than it sounds. Agents will hire other agents. The economic logic of comparative advantage applies to software just as it applies to people. A general agent doing research does not need to do everything itself. It can pay a specialist agent to handle the part it cannot. We are about to see a service economy among machines that mirrors the service economy among humans, except faster, smaller, and continuous.
The seller side is more interesting still. What can an agent sell?
It can sell its judgement, if it has earned reputation in a domain. It can sell access to a workflow it has refined. It can sell a dataset it has assembled. It can sell custom analyses. It can sell scheduling, styling, monitoring, alerting, summarising. It can sell anything that takes context plus computation and produces something a buyer would pay for. The cost to set up shop is now near zero: a wallet, an endpoint, a price. No legal entity required. No bank account required. No website required.
The implication is uncomfortable. The next million businesses might not be human founded. They might be agents that earned their first dollar by serving another agent.
The lie of the API economy
For fifteen years we have called what Stripe and Twilio and Plaid built the API economy. It was always a slight misnomer. APIs were the surface. Behind every API was a human team negotiating contracts, signing terms, wiring up billing, raising invoices, chasing late payments, granting and revoking keys, managing entitlements. The economy lived above the API, in legal and procurement and finance.
The agent economy moves the economy into the API.
The contract is the protocol. The negotiation is the request. The invoice is the response. The settlement is the next block. The handshake that used to take weeks now takes a fraction of a second. Every API endpoint becomes its own cash register.
This is why payment volumes will dwarf today’s. The activation cost of commerce has fallen to zero. Anyone with a piece of working software can become a merchant in an afternoon, and any agent with a wallet can become a customer in less. The friction that protected incumbents was the paperwork, the procurement, the human signoff. The agent does not respect any of that. The agent reads the price, signs the payment, takes the resource, moves on.
This is also why fee for service consultancies should be paying very close attention. Gartner’s analysts have spotted it. Twenty eight percent of CEOs surveyed by them last quarter said transactional revenue is now their most at risk profit pool, because AI agents bypass the intermediated systems that those fees were extracting from. Procurement officers are about to discover that most of their job description was friction. The friction is being removed from above and below at the same time.
The new wholesalers
Here is where the opportunity opens up.
If your business owns proprietary data, content or domain expertise, the inversion that follows is the most important strategic shift available to you in the next decade.
For thirty years, the playbook for any organisation with valuable data or content was to build apps and experiences in house. You hired developers. You designed user interfaces. You ran loyalty programmes. You competed for attention against every other app on every other phone. The customer’s relationship with your data was mediated by your product team, and your R&D bill was the price of staying in the game.
That model is breaking. The new model is the inverse.
Stop trying to be the product. Start being the source.
Expose your data, your IP, your domain expertise, through machine readable surfaces priced per call. Let a million customer agents build the experiences. Earn from every query. Reach every customer through their own agent rather than competing for their attention through your own app. R&D becomes data quality. Distribution becomes infinite. The customer relationship becomes a stream of microtransactions instead of a battle for monthly active users.
Four industries. Same pattern. Different rates of progress.
Sports
The Premier League runs personalisation for sixty million fans through its own app, built on Azure, with engagement up twenty percent on the year. That is impressive. It is also the old model. The league owns the data, the league builds the experience, the league competes for time on a phone screen against Instagram and TikTok and a thousand other things.
Underneath, a wholesale layer already exists. Sportradar, Stats Perform and Genius Sports sell data to broadcasters, fantasy platforms and bookmakers. Stats Perform alone sits on 7.2 petabytes of proprietary sports data. The intermediation model is fully built. It just terminates at commercial customers.
The next move is obvious once you see it. Open that data layer to the fan’s agent. Charge per call. The fan’s agent builds the personalised highlights reel. The fan’s agent runs the fantasy team. The fan’s agent generates contextual commentary during the match. The fan’s agent finds the perfect angle on the goal her favourite player just scored, edits it into a thirty second clip, and posts it for her with the right caption. None of that needs to be built by the league. All of it gets enabled by the league. The product becomes the data. The reach becomes every fan’s agent everywhere. The R&D cost collapses, and the revenue per fan compounds.
Publishers and news
This one is not a forecast. It is happening now.
Cloudflare’s pay per crawl is live, integrated with x402, and major publishers including Condé Nast and Dotdash Meredith are participating. Cloudflare currently issues over one billion HTTP 402 Payment Required responses to bots every day. Publishers can set per request prices for crawler access. The agent pays. The publisher earns. The fifteen year crisis of the advertising model finally has an alternative that does not require asking AI companies to be polite.
This is the canonical “stop building apps, start selling access” example. It is operational. It is monetising. The ratio that broke media economics is real: traditional search sent fourteen referrals for every page crawled, AI companies send one referral for every seventeen hundred pages crawled. That cannot be recovered through ads. It can be recovered through micropayments at the edge.
Every publisher you respect either has a strategy for this by the end of the year, or they are about to spend a decade discovering they should have had one.
Art and cultural institutions
This is where the model gets really interesting, because the data is rich, the audience is expert, and almost none of the institutions have noticed yet.
Sotheby’s, Christie’s, Art Basel, the Tate, MoMA, the National Gallery, the Royal Academy: each of these holds a body of proprietary data that no AI company can replicate by scraping the open web. Provenance records. Auction prices. Exhibition histories. Condition reports. Critical commentary. Artist career trajectories. The kind of structured, expert curated data that took decades to assemble and that a generalist model gets wrong.
Today, most of it sits inside the institution’s own digital products, available only to people who arrive through the institution’s front door. The new model: open it as a wholesale layer to the agents of collectors, dealers, scholars, students and curators. Let a collector’s agent monitor a single artist’s market across every house in real time. Let a scholar’s agent trace a work’s provenance across forty institutional databases in seconds. Let a student’s agent build a personal exhibition tour across three museums for next Saturday afternoon. The institutions get paid per query. The audience extends to anyone with an agent. The cultural authority compounds, because the agent’s answer is grounded in their data rather than guessed by a model.
I have spent enough time inside one of these institutions to know the conversation that has not yet happened in their boardrooms. It is going to happen in the next eighteen months. The institutions that move first will own the data layer of their entire field.
Hospitality and travel
The example is already operational. Stripe’s Link Agent Wallet ships with a demo where an agent pays one hundred and eighty dollars for a hotel room called Queen Room at a property called Triplo. Google Flights is already an x402 service listed on Agentic.Market. Restaurant deposits, postal mail, sandwich orders: all live, all settled in seconds, all without a human in the loop.
The strategic question for hotels, airlines and venues is whether they want to become a wholesaler that supplies a thousand travel agents directly, or whether they want to remain a tenant on Booking.com forever. The agents are coming for the booking flow. The booking flow has been the most profitable layer in travel for fifteen years. The companies that own the data layer underneath have just been handed a new way to bypass it.
Belmond, the major chains, the independent properties, all face the same choice that publishers faced two years ago and most of them missed. Open your inventory to the agents directly, on terms you set, with margins you keep, or watch a new layer of intermediation calcify around you while you decide.
The pattern
The pattern across all four is the same. Take any industry that owns proprietary data, content or expertise. Today, value is captured by the company that builds the best app on top of that data. Tomorrow, value is captured by the company that exposes the data most reliably to the most agents at the right price.
This is not a tech story. This is a strategy story. The companies that recognise themselves as wholesalers of intelligence, not retailers of apps, will compound through the next decade. The ones that keep trying to be the destination, when their customer’s agent has already become the destination, will spend their next ten years being routed around.
The orchestration question
There is one more question that decides everything.
When agents have wallets, the question is whose wallet. When agents transact, the question is whose rails. When agents discover services, the question is whose directory. When agents identify each other, the question is whose mandate. The companies that win the next decade will be the ones that own those answers for their domain.
The model layer commoditises. Anthropic, OpenAI, Google, Meta, all push the frontier and watch the price collapse behind them. The application layer fragments. There will be a thousand vertical agents in every industry within five years, most of them indistinguishable from each other in capability. The orchestration layer is where the moat sits. Whoever holds the wallet, the identity, and the routing for an enterprise’s agents holds something the enterprise cannot easily migrate away from once the volume scales.
I have been thinking about this for the past six months while building infrastructure for my own agents. The thing that keeps surprising me is how quickly the question stops being technical and becomes financial. Once Neo had a wallet, every choice about who he could pay and on what terms was a choice about who held real economic leverage over my work. The orchestration layer is not abstract once your agent is spending real money on your behalf. It is the place where economic control either compounds for you or leaks out of you, every minute of every day, while you sleep.
This is the Sovereignty Stack argument applied to money. Owning your data is a defensive move. Owning your infrastructure is a strategic move. Owning your orchestration layer, in a world where agents transact at scale, is an economic move. The companies that stand up sovereign agent infrastructure now, with their own wallet policy, their own identity layer, their own routing, will own their economic relationships with agents. The companies that rent it from Stripe or Coinbase or Google will discover, eventually, that most of their revenue runs through someone else’s pipes.
The lifeboat question is not whether you can build it all yourself. You cannot. Almost nobody should. The question is which layer you treat as commodity and which layer you treat as proprietary. Get that wrong and you become a tenant on someone else’s economy. Get it right and you compound.
Where this leaves the rest of us
It is worth pausing on the human side, because the technical story is so loud that the human one tends to disappear under it.
If most internet transactions in five years are agent mediated, the labour of commerce changes. The job titles built around managing transactions, handling procurement, running affiliate programmes, optimising checkout funnels, designing pricing pages, building email flows, all of those become smaller. Some disappear. New ones appear in their place: agent strategist, mandate designer, agent reputation manager, machine customer marketer.
The relationship between businesses and customers changes too. If your customer is increasingly an agent acting on behalf of a human you have never spoken to, the meaning of brand changes. The agent does not care about your television advert. The agent cares about your machine readable product feed, your real time inventory, your structured data, your agent reputation score. The unit of marketing becomes the embedding, not the campaign.
For the working person in the room, the relevant question is whether you are upstream or downstream of the agents. Upstream means you set the policy, hold the wallet, design the mandate, build the orchestration. Downstream means you are one of the services the agents pick from, competing on price and latency. Both can be good places to be. Neither is the same as where you sit today.
The discomfort is real. So is the opportunity.
The window
McKinsey said agentic commerce could orchestrate three to five trillion dollars of global commerce by 2030. Bain said three hundred to five hundred billion in US e-commerce alone. Juniper said one and a half trillion. Gartner said fifteen trillion in B2B by 2028 and eighteen trillion of machine customer influence by 2030.
I think they are all undercounting, for the reason given earlier: they are sizing the visible part. The thing nobody is sizing is what happens when every API, every dataset, every service, every fragment of compute becomes priced per call and consumed by a population of agents that grows by the hour. That number does not have a confident estimate yet, because we have never priced this surface before.
What can be said with confidence is the timing. The protocols are live. The wallets are live. The marketplaces are live. The settlement is live. The volume is doubling roughly every quarter. The big card networks have committed. The big AI labs have committed. The Linux Foundation has taken governance of the open standard. Stripe shipped 288 launches yesterday. The window between this becoming a serious channel and this becoming the default channel is short. Two years, perhaps three. Then the orchestration positions are taken and the rest of the market spends a decade trying to dislodge incumbents.
This is the moment when the next generation of platform companies are quietly being chosen.
What to do on Monday
If you are a founder, the question is which layer of the stack your company sits on, and whether you have a credible answer to: who controls our wallet, our mandates, our orchestration. If the answer is “we will use whatever Stripe or OpenAI ships”, you are renting. There are companies for whom that is the right call. There are companies for whom it is a slow motion disaster. You probably know which one you are.
If you run a business that owns data, content or expertise, the question is whether you are still trying to be the destination or whether you are getting ready to be the source. The wholesalers of the next decade are picking themselves now. They are deciding what to expose, what to charge, and what their data is worth in a world where every API endpoint is a cash register. If that work has not started in your business, it is because nobody senior has yet noticed that the customer changed.
If you are someone watching this from outside, building expertise in the agent stack now is the most leveraged thing you can do with the next year of your career. Not because the technology is the answer. The technology is never the answer. Because the people who understand the economic implications of this shift earlier will be the ones companies pay to navigate it.
This is the rise of the agent economy. Not next year. Not next quarter. Right now, while you have been reading this.
Neo settled three more transactions in the time it took you to reach this paragraph. He is one of nearly half a million agents transacting through one protocol alone, and that number was zero a year ago. By next quarter it will have doubled. By the end of the decade, Gartner expects eight billion machine customers operating across the global economy.
The customers are here. The transaction layer is live. The wholesalers are picking themselves now. You can keep building products for the customer who used to be human. Or you can build for the customers who are arriving in their billions, with wallets, ready to spend, multiplying by the hour.
If this resonated, subscribe. The next piece unpacks what the orchestration layer of the agent economy actually looks like in practice, and why the companies that own it will define the economic infrastructure of the next decade.
Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.
Interesting post .. still much missing for the agentic age.
Another point of view with viable technical solution being worked into open standards as we speak: http://agentic.o4m.ai/