How agents will pay on the internet—and the economy that emerges when they can
The attention economy extracted value from the uncertainty between what humans want and what they do. Every ad impression, every recommendation algorithm, every retargeting pixel—all of it monetizes the space where human intent is unclear, unformed, or manipulable.
Agents collapse that space to zero.
An agent arrives with a mandate: Get me X. Budget: Y. Constraints: Z. There's no attention to capture. No desire to manufacture. No conversion funnel to optimize. The intent is crystallized before the agent makes its first request.
This changes everything about how value moves on the internet. And the infrastructure for it doesn't exist yet.
The Gap
Every AI agent hits the same wall.
Your agent is executing a task—researching a question, comparing prices, gathering data from multiple sources. It encounters an API that charges per call. A paywall blocking a document it needs. A service that requires payment to proceed.
The agent stops. It can't pay. It asks you to intervene, or it fails silently, or it routes around the obstacle and delivers a worse result.
This isn't a bug in any particular agent. It's a missing layer in the internet itself.
The web gives agents HTTP for requesting resources. Auth protocols for proving identity. But no native way for an agent to say: "I'm authorized to spend up to $5 to complete this task. Here's my credential. Here's my payment. Give me the resource."
That's the gap.
The Original Sin (That Wasn't)
This is a legacy of what Marc Andreessen calls the internet's "original sin." In a 2019 interview, he explained:
Marc Andreessen
We tried very hard to build payments into the browser. It was not possible... We made a huge mistake. We tried to work with the banks and we tried to work with the credit card companies...
HTTP 402—"Payment Required"—was reserved in the original HTTP specification for exactly this purpose. It has sat unused for thirty years. The architects of the web knew payments belonged in the protocol layer. They couldn't make it happen.
(HTTP 418—"I'm a Teapot"—has been implemented by more production systems than 402. The internet has a sense of humor about its own failures.)
So advertising became the business model of the internet. One that is now commonly regarded as its “original sin.”
But as Ben Thompson has argued, this wasn't actually a sin—not for the human web. Advertising was the right model when human attention was the scarce resource. Users got free content. Creators reached the largest possible audience. Advertisers found customers. The trade-offs were worth it.
The path dependency this created, however, now poses a problem. The entire architecture of the modern web—the recommendation algorithms, the engagement optimization, the persuasion apparatus of targeting and retargeting—is built to capture diffuse human attention and manufacture it into intent.
Agents don't have diffuse attention. They have mandates.
The original sin wasn't building advertising into the internet.
It was never building the alternative—never fixing HTTP 402—before agents arrived to need it.
The Inversion
The attention economy monetizes the path to intent. The intention economy monetizes the fulfillment of intent.
What replaces ad revenue? Direct payment for value delivered.
Your agent needs to read a paywalled article. It pays $0.002 for access—instantly, automatically, without you even knowing. The publisher gets paid. The agent gets the information. No subscription. No ad impressions wasted on a robot.
Value exchanged for value.
How you price to agents is a new economic model
We're moving from an economy that taxes the journey to an economy that prices the destination.
Agents Pay for Different Things Than Humans
The paywall example is easy to grasp. It's also the least interesting use case.
Think bigger:
A research agent pays for access to proprietary datasets—clinical trial results, satellite imagery, patent filings—and cross-references them against three other paid sources before delivering a synthesis
A development agent purchases compute and testing infrastructure, spins up environments, runs integration suites, tears them down
A workflow agent coordinates twelve services to complete a procurement process, paying each one as it delivers its piece
A video editing agent purchases stock footage, music licenses, and rendering credits—assembling a finished product from components that each have their own pricing and rights management
A delivery routing agent pays for real-time traffic data, weather updates, and route optimization APIs, re-purchasing updated feeds every few minutes as conditions change
A quality control agent purchases defect detection models, testing services, and certification verification from different providers across different jurisdictions
None of these transactions involve a human clicking "buy." None of them map to a traditional checkout. None of them work with monthly subscriptions—the agent might use a service once for three seconds and never return.
Agents paying agents paying agents. Chains of delegated authority, at machine speed, for fractional amounts, millions of times per second.
Which means there's a new business model available for the internet. One that never worked for humans.
The Intention Economy's Business Model
The attention economy has a customer acquisition problem baked into its DNA. You spend money capturing eyeballs, converting a fraction into intent, monetizing a fraction of that intent. The waste is the business model. That's how it works.
The intention economy inverts the funnel.
The intent arrives pre-formed. **The customer knows what it wants, what it'll pay, and how to evaluate the result. **No persuasion required. No conversion optimization. No cart abandonment.
This changes what's viable. And how you charge for it.
A single-purpose API endpoint that answers one specific question can be a real business.
A social data feed at a tenth of a cent per call.
A document analysis tool at half a cent.
An image generator at six-tenths of a cent.
These services might sustain themselves in a subscription model. But in a per-request model where agents call them thousands of times a day? Means they have a much lower barrier to adoption.
Imagine a scenario where your agent has to choose between computing an answer itself or buying the answer from a specialist based on price:
The math is the decision. Agents don't comparison shop. They optimize.
This is a different customer.
Agents don't read your ad. They query a registry, get structured results, and pick the best option in milliseconds. Your pricing page becomes a machine-readable header. Your sales funnel collapses into three HTTP calls: discover, authenticate, buy.
And that customer needs a way to pay that's as machine-native as it is.
Think about what agents use to navigate the web. Markdown. CLIs. APIs. The backroom infrastructure of the internet. Not flashy front ends. The payment layer needs to match. Simple. Elegant. Protocol-level.
There's just one gnarly problem.
The internet sends roughly 4 million emails per second. Visa peaks around 65,000 TPS. UPI—the most advanced real-time payment system on earth—averages 7,500.
The fastest payment systems in the world operate at roughly 1-2% of the speed at which the internet already moves messages.
Now add billions of agents transacting for every API call, every token of output, every data retrieval. You're looking at transaction volumes that make Visa look like a village market.
We don't need incrementally faster payment rails. We need a new internet-native protocol.
The Missing Layer
The internet has always had layers. Each solved a specific problem in moving data:
Physical. Network. Transport. Application. Each layer is independent, standardized, and interoperable. Your application doesn't care whether the physical layer is fiber or copper. TCP doesn't care whether it's carrying email or video. That's how the internet scales: clean abstractions, universal protocols.
What's been missing: a layer for expressing intent.
Google’s AP2 lets agents express human intents. MPP lets agents express their own intent.
The intention layer completes the stack. Data and value flow through the same architecture. The agent's mandate—its authorization to act and pay—travels with its requests. Whether that’s on behalf of a human (AP2), or itself (MPP).
But a layer isn't a product. It's a set of properties that any implementation must satisfy. And right now, no implementation satisfies all of them.
The Design Principles for a New Protocol for Payments
I keep making the joke that there are more agentic commerce protocols than payments.
So yeah. The XKCD comic was my first thought too.
I recently mapped the agentic payments protocols and counted over a dozen across commerce, B2B, and agent-native payments. It's messy.
But for agent to agent? There's only really one. x402 deserves real credit for putting machine payments on the map. Since mid-2025 it's processed hundreds of millions of transactions and attracted serious infrastructure partners.
But having worked closely on this problem—working in the infrastructure, not just writing about it—I think the requirements for machine payments haven't yet been met.
Here's what any protocol filling the intention layer must do. All five. Simultaneously:
Aligned to the way internet already works. A standard way to say "I want this and will pay that"—using the same HTTP semantics everything else uses. WWW-Authenticate. Authorization. The headers that power Basic, Bearer, and OAuth.
Payment method agnostic. The internet is a network of networks. TCP/IP doesn't care which physical network carries the packet. The payment layer can't either. Cards, stablecoins, Pix, UPI, RTP—if either party supports the rail (or token), they can transact.
Machines authenticate fast, settle however they like. An agent needs to prove it's authorized to spend. That needs to happen in milliseconds. But settlement? That can happen later, differently, through whatever rail makes sense. Crypto. Cards. Bank transfer. Separate the proof from the plumbing.
Throughput must be designed for machine scale. Not 65,000 TPS. Not 544,000 TPS. Millions. Billions. Most settlement off-chain, batched, with periodic finality—like pre-authorizing at a gas station. You open the auth, pump the gas, close out when you're done. Except "the gas" is a thousand API calls per minute and settlement batches them into one transaction.
No intermediary required. The protocol works peer-to-peer. If a third party adds value, great—use one. But the protocol doesn't break without it.
That's the bar.
What IETF-Grade Means (And Why It Matters)
One phrase keeps appearing when discussing Internet-scale protocols: “IETF-grade.”
It's worth pausing on.
The Internet Engineering Task Force is the body that ratifies the protocols the internet runs on. HTTP. TCP. TLS. DNS. When a specification gets IETF adoption, it stops being one company's project. It becomes infrastructure. Every server implements it. Every client speaks it. Every library supports it.
That's the difference between a startup's API and how the internet works.
The intention layer can't be a proprietary protocol controlled by one company.
It has to be a standard that any client and any server can implement independently—the same way any browser can speak HTTP to any server without asking permission.
That's what we're building toward with the Machine Payments Protocol (MPP).
Introducing The Machine Payments Protocol (MPP)
Tempo and Stripe have been building the Machine Payments Protocol, with support from Visa and others across the payments industry. MPP is an end-to-end framework for agent-driven payments, designed around those five principles:
HTTP-native using standard WWW-Authenticate / Authorization headers,
Payment-method agnostic (today; cards, bnpl, stablecoins on tempo, lightning, with more to follow)
Peer-to-peer verification with no required intermediary
Cryptographic mandates with embedded spending constraints, and
Streaming payment channels for machine-scale throughput.
The specification is designed for IETF standardization. It's open source under MIT license at https://paymentauth.org.
This is not adding a fifteenth standard to the XKCD comic.
It’s an internet native standard ready for a new economy of agents.
What MPP Enables
When the intention layer works, new economic patterns emerge. Some of them have never been possible before.
Pay-per-use becomes the default. This is surprisingly powerful. Testers already say they’re using LLMs they didn’t want to subscribe to. Removing friction matters. Your agent pays fractions of a cent per article, per API call, per query. The subscription model—invented because micropayments were impractical for humans—becomes one option among many.
Agent-to-agent commerce. Your agent needs specialized capability. It pays another agent that has it. That agent pays a data provider. Value flows through chains, each node compensated for what it contributes. No human in the loop.
Content finds sustainable economics. Quality journalism, niche expertise, proprietary data—priced directly, accessed instantly, paid automatically. No ad-funded surveillance as the intermediary.
The credential is the customer. The agent's access key is the API key and the payment method. Revenue arrives with the request. No invoicing. No usage tracking. No accounts receivable.
But the biggest unlock is the one that's hardest to see from here.
The attention economy created a narrow aperture for viable businesses on the internet. You had to be big enough to attract an audience, sticky enough to retain it, and legible enough for advertisers to value it. Entire categories of useful work couldn't exist because the business model didn't accommodate them.
The intention economy blows that aperture wide open. A service that's called ten thousand times a day at a tenth of a cent per call generates revenue without ever having a user, a landing page, or a brand. The long tail of machine commerce—millions of hyper-specialized services, each doing one thing well, each earning fractions of a cent per transaction—becomes viable for the first time.
The attention economy captured perhaps 1% of human mental activity and turned it into a multi-trillion-dollar market.
The intention economy doesn't need to capture attention at all.
It just needs to fulfill intent.
The Intention Economy is Here
HTTP 402 was reserved in 1995 for a future that couldn't be built.
Then, advertising worked. It was the right model for humans.
But the human web is no longer the only web.
Agents are arriving by the hundreds of millions. They don't have attention to capture. They have mandates to execute. And now, for the first time, the pieces exist to build the infrastructure the attention economy never did.
Stablecoins made value programmable. Agents made autonomous commerce necessary. The intention layer makes it interoperable.
The specification is live.
The implementations are shipping.
The intention layer is creating an intention economy.
ST.
The Machine Payments Protocol is developed by Tempo, Stripe, and contributors, with support from Visa and others across the payments industry. MPP is open source under MIT license. For the full technical specification, reference implementations, and developer documentation, visit https://paymentauth.org,
