Simon Taylor published The Intention Layer. This time he goes beyond simply mapping the agentic payments landscape, addressing at the protocol level why the internet never had a payment layer and why it needs one now. His answer: the Machine Payments Protocol (MPP), being built by Tempo and Stripe.
Taylor's core argument is the shift from the attention economy to the intention economy.
The ad-driven internet monetizes the space between what humans want and what they do. Every ad impression, recommendation algorithm, and retargeting pixel makes money at the point where human intent is unclear or unformed.
Agents collapse that space to zero. "Buy me X. Budget: Y. Constraints: Z." There is no attention to capture, no desire to manipulate, no conversion funnel to optimize. Intent is crystallized before the first request is ever made.
The key metaphor for this shift: moving from "an economy that taxes the journey" to "an economy that prices the destination."
The attention economy first captures eyeballs, converts a fraction into intent, and monetizes a fraction of that intent.
The intention economy inverts the funnel. Intent arrives pre-formed. The customer already knows what it wants, what it will pay, and how to evaluate the result. No persuasion needed, no conversion optimization, no cart abandonment.
But there is a problem. The internet has no native way for an agent to say: "I'm authorized to spend up to $5 to complete this task. Here are my credentials, here is my payment method, give me the resource." HTTP 402 "Payment Required" was reserved in the original HTTP spec in 1995, but has never been implemented in thirty years.
The scale problem Taylor presents is also striking. The internet sends roughly 4 million emails per second. Visa's peak throughput is about 65,000 TPS. UPI, the most advanced real-time payment system in the world, averages 7,500 TPS. The fastest payment systems operate at 1–2% of the speed the internet already moves messages. Add billions of agents transacting on every API call, every token output, every data retrieval, and Visa starts looking small.
MPP aims to solve this with five principles:
Just as TCP/IP doesn't care which physical network carries the packet, the payment layer must accept any rail—cards, stablecoins, BNPL, Lightning Network, UPI, RTP.
It uses existing WWW-Authenticate and Authorization headers—the same mechanism that powers Basic, Bearer, and OAuth.
Proving payment authorization happens in milliseconds, but settlement happens later, through whatever rail is appropriate.
Not 65,000 TPS, not 544,000 TPS—millions to billions of transactions. Most settlement is batched off-chain with periodic finality.
The protocol works peer-to-peer. Third parties can add value if they want, but the protocol doesn't break without them.
The biggest change the intention economy opens up is the liberation of the long tail.
The attention economy narrowed the range of viable businesses. You had to be big enough to attract an audience, addictive enough to retain them, and valuable enough for advertisers to care. In the intention economy, a service called 10,000 times a day at $0.001 per call generates revenue without a single user, landing page, or brand. Millions of hyper-specialized services, each doing one thing well and earning fractions of a cent per transaction, become viable for the first time.
The cards vs. stablecoins debate over payment methods for agentic commerce is heating up. @zephyrlogs broke down this debate on Twitter, and the key point is that "giving cards to agents" isn't one design—it refers to several distinct control models. The differences matter in practice.
Why do cards show up first?
Cards are already integrated into most of the merchant internet. Merchants accept them, checkouts are built around them, and fraud detection, dispute handling, and refund processes all assume cards. Even systems designed specifically for AI agents often start at or stay close to card rails in their payment layer.
The four models are as follows.
Shared Card Access: The most familiar model. The agent uses an existing personal or company card. The advantage is simplicity—less setup, fewer moving parts, lower initial friction. The tradeoff is scope. The credential is broad. When multiple agents or workflows share it, spending is hard to separate cleanly, limits apply at the account level, and it's difficult to trace which agent initiated which purchase.
Dedicated Virtual Card: A new card is issued for a narrower unit. This model is already familiar outside of agentic commerce—Ramp and Brex use it for corporate spend, Privacy.com uses the same logic for consumers. Instead of giving an agent a broad payment credential, you issue one designed for the specific task. Narrower scope means clearer attribution and easier post-hoc review.
Tokenized Card Delegation: The agent never receives raw card credentials. Instead it operates through a delegated token or network-backed credential. Mastercard Agent Pay, Visa Intelligent Commerce, and Stripe Shared Payment Tokens are heading this direction. A virtual card, however narrowly scoped, is still a card credential. The tokenized model pushes delegation logic into the token layer itself. This is also why the current debate gets confusing—some newer systems are discussed as alternatives to card-based commerce, when they're more accurately understood as new delegation layers on top of card-based commerce.
Hybrid Models: Most real systems won't fit neatly into one bucket. A platform might hold funds in a single account while generating separate virtual cards per agent action; a buyer approves a payment method once and then the agent uses tokenized delegated access; or card credentials are used for merchant acceptance while policy is enforced in a software layer before authorization.
Ultimately, the core question isn't "cards or not." It's where does the funding live, where does the credential live, and where does the policy live. How these three separate determines the system's risk profile.
Cards fit naturally in agent-to-business commerce. In that environment, they bring existing merchant acceptance, familiar checkout behavior, and mature dispute and refund processes.
Where cards don't fit is agent-to-agent environments—no browser, no consumer-style purchase flow, no chargebacks, no merchant acquiring stack. Micro-amount, high-frequency payments for API calls, compute, and content access also clash with the card model's transaction structure.
Gabriel Shapiro introduced the concept of "Slop Tokenization" on Twitter. His core claim is that most tokens represent nothing in an architectural sense.
The blockchain performs no essential function in the relationship between the token and the thing it claims to represent.
Remove the blockchain and the underlying economic or legal reality is unchanged.
The token is ornamentation on a structure that doesn't need it.
Shapiro first defines three forms of "clean tokenization."
System-intrinsic tokens: ETH and BTC are the paradigmatic cases. Native elements of the protocol, with rights enforced by consensus rules and no off-chain referent. The chain doesn't represent something—it IS the thing. ETH is Ethereum's access credential, enforced by code every node runs. Remove ETH and the system ceases to function.
Contractually operative tokens: USDC is the clearest example. Circle's Terms of Service make holding USDC the mechanism for exercising redemption rights. The token isn't the dollar, but holding the token activates your right to get dollars. A binding contract makes token-holding the trigger for enforceable claims against the issuer.
Legally constitutive tokens: An on-chain stock ledger is the paradigmatic case. If a corporation's governing documents designate the on-chain state as the authoritative record of holder rights, the blockchain doesn't track ownership—it constitutes ownership.
According to Shapiro, if none of these three patterns apply, it's slop. The one-line test: is the token merely a UI pointer to an off-chain registry that can ignore you when inconvenient?
The case studies Shapiro presents are overwhelming.
BitClout scraped 15,000 Twitter profiles without consent and attached bonding-curve "creator coins" to each. No enforceable right was ever granted to token holders.
Rally raised $57 million in VC funding to build a creator coin platform, then abruptly shut down its sidechain in January 2023 citing "macro headwinds." The blockchain protected nothing. The things of value—the creator's willingness to provide benefits, a functioning platform—had never existed on-chain in the first place.
The Moonbirds case is peculiar. Minted in April 2022 at 2.5 ETH, reaching a floor of 30 ETH, with holders promised commercial IP rights. Then in August of the same year, Kevin Rose unilaterally switched to a CC0 (Creative Commons Zero) license. Anyone—holder or not—could use any Moonbird image for any purpose, making the legal standing of someone who paid 30 ETH identical to someone who paid nothing. The key lesson isn't the switch itself, but that the switch was possible at all. IP rights had never been embedded in the token in any legally operative way. The issuer's ToS carried the rights, and the ToS could be changed without holder consent.
BAYC is somewhat better. Yuga Labs' terms state that "when you purchase an NFT, you own the underlying Bored Ape, the Art, completely," linking an unlimited commercial use license to cryptographic ownership verification. Holders have launched restaurants, seltzer brands, and TV shows. The concession that "at no point may we seize, freeze, or otherwise modify the ownership of any Bored Ape" is also key. It isn't perfect, but a binding legal agreement making token-holding the trigger for enforceable commercial rights puts it in a fundamentally different category from Moonbirds.
Nasdaq's September 2025 proposed rule change is also mentioned—settling equity trades in "tokenized form," except the blockchain doesn't determine price, match orders, settle trades, or hold assets. It receives a notification after the fact from the same centralized systems. The press will write "Nasdaq brings stock trading on-chain," but the reality is a notification layer for a system that functions identically without it.
BlackRock's BUIDL is a somewhat better case. There is real utility: 24/7 transfers, programmable collateral, collateral acceptance on Deribit and Binance. But the underlying assets are held by BNY Mellon, managed by BlackRock, and if Securitize disappears, the token doesn't self-execute a bond redemption.
Shapiro's argument gets genuinely scary when he points out that this problem extends beyond individual tokens to threaten blockchains themselves. Take Ethereum as an example: ETH itself is the paradigm of clean tokenization, but for ETH to be valuable, Ethereum must be useful, and for Ethereum to be useful, meaningful assets and contracts must actually settle there. If everything on the chain other than ETH is merely "the real thing lives off-chain, the token just points to it"—then what? Ethereum is just the example here, but the broader point stands: from this perspective, a blockchain becomes little more than a censorship-resistant archive of photocopies.
Simon Taylor published The Intention Layer. This time he goes beyond simply mapping the agentic payments landscape, addressing at the protocol level why the internet never had a payment layer and why it needs one now. His answer: the Machine Payments Protocol (MPP), being built by Tempo and Stripe.
Taylor's core argument is the shift from the attention economy to the intention economy.
The ad-driven internet monetizes the space between what humans want and what they do. Every ad impression, recommendation algorithm, and retargeting pixel makes money at the point where human intent is unclear or unformed.
Agents collapse that space to zero. "Buy me X. Budget: Y. Constraints: Z." There is no attention to capture, no desire to manipulate, no conversion funnel to optimize. Intent is crystallized before the first request is ever made.
The key metaphor for this shift: moving from "an economy that taxes the journey" to "an economy that prices the destination."
The attention economy first captures eyeballs, converts a fraction into intent, and monetizes a fraction of that intent.
The intention economy inverts the funnel. Intent arrives pre-formed. The customer already knows what it wants, what it will pay, and how to evaluate the result. No persuasion needed, no conversion optimization, no cart abandonment.
But there is a problem. The internet has no native way for an agent to say: "I'm authorized to spend up to $5 to complete this task. Here are my credentials, here is my payment method, give me the resource." HTTP 402 "Payment Required" was reserved in the original HTTP spec in 1995, but has never been implemented in thirty years.
The scale problem Taylor presents is also striking. The internet sends roughly 4 million emails per second. Visa's peak throughput is about 65,000 TPS. UPI, the most advanced real-time payment system in the world, averages 7,500 TPS. The fastest payment systems operate at 1–2% of the speed the internet already moves messages. Add billions of agents transacting on every API call, every token output, every data retrieval, and Visa starts looking small.
MPP aims to solve this with five principles:
Just as TCP/IP doesn't care which physical network carries the packet, the payment layer must accept any rail—cards, stablecoins, BNPL, Lightning Network, UPI, RTP.
It uses existing WWW-Authenticate and Authorization headers—the same mechanism that powers Basic, Bearer, and OAuth.
Proving payment authorization happens in milliseconds, but settlement happens later, through whatever rail is appropriate.
Not 65,000 TPS, not 544,000 TPS—millions to billions of transactions. Most settlement is batched off-chain with periodic finality.
The protocol works peer-to-peer. Third parties can add value if they want, but the protocol doesn't break without them.
The biggest change the intention economy opens up is the liberation of the long tail.
The attention economy narrowed the range of viable businesses. You had to be big enough to attract an audience, addictive enough to retain them, and valuable enough for advertisers to care. In the intention economy, a service called 10,000 times a day at $0.001 per call generates revenue without a single user, landing page, or brand. Millions of hyper-specialized services, each doing one thing well and earning fractions of a cent per transaction, become viable for the first time.
The cards vs. stablecoins debate over payment methods for agentic commerce is heating up. @zephyrlogs broke down this debate on Twitter, and the key point is that "giving cards to agents" isn't one design—it refers to several distinct control models. The differences matter in practice.
Why do cards show up first?
Cards are already integrated into most of the merchant internet. Merchants accept them, checkouts are built around them, and fraud detection, dispute handling, and refund processes all assume cards. Even systems designed specifically for AI agents often start at or stay close to card rails in their payment layer.
The four models are as follows.
Shared Card Access: The most familiar model. The agent uses an existing personal or company card. The advantage is simplicity—less setup, fewer moving parts, lower initial friction. The tradeoff is scope. The credential is broad. When multiple agents or workflows share it, spending is hard to separate cleanly, limits apply at the account level, and it's difficult to trace which agent initiated which purchase.
Dedicated Virtual Card: A new card is issued for a narrower unit. This model is already familiar outside of agentic commerce—Ramp and Brex use it for corporate spend, Privacy.com uses the same logic for consumers. Instead of giving an agent a broad payment credential, you issue one designed for the specific task. Narrower scope means clearer attribution and easier post-hoc review.
Tokenized Card Delegation: The agent never receives raw card credentials. Instead it operates through a delegated token or network-backed credential. Mastercard Agent Pay, Visa Intelligent Commerce, and Stripe Shared Payment Tokens are heading this direction. A virtual card, however narrowly scoped, is still a card credential. The tokenized model pushes delegation logic into the token layer itself. This is also why the current debate gets confusing—some newer systems are discussed as alternatives to card-based commerce, when they're more accurately understood as new delegation layers on top of card-based commerce.
Hybrid Models: Most real systems won't fit neatly into one bucket. A platform might hold funds in a single account while generating separate virtual cards per agent action; a buyer approves a payment method once and then the agent uses tokenized delegated access; or card credentials are used for merchant acceptance while policy is enforced in a software layer before authorization.
Ultimately, the core question isn't "cards or not." It's where does the funding live, where does the credential live, and where does the policy live. How these three separate determines the system's risk profile.
Cards fit naturally in agent-to-business commerce. In that environment, they bring existing merchant acceptance, familiar checkout behavior, and mature dispute and refund processes.
Where cards don't fit is agent-to-agent environments—no browser, no consumer-style purchase flow, no chargebacks, no merchant acquiring stack. Micro-amount, high-frequency payments for API calls, compute, and content access also clash with the card model's transaction structure.
Gabriel Shapiro introduced the concept of "Slop Tokenization" on Twitter. His core claim is that most tokens represent nothing in an architectural sense.
The blockchain performs no essential function in the relationship between the token and the thing it claims to represent.
Remove the blockchain and the underlying economic or legal reality is unchanged.
The token is ornamentation on a structure that doesn't need it.
Shapiro first defines three forms of "clean tokenization."
System-intrinsic tokens: ETH and BTC are the paradigmatic cases. Native elements of the protocol, with rights enforced by consensus rules and no off-chain referent. The chain doesn't represent something—it IS the thing. ETH is Ethereum's access credential, enforced by code every node runs. Remove ETH and the system ceases to function.
Contractually operative tokens: USDC is the clearest example. Circle's Terms of Service make holding USDC the mechanism for exercising redemption rights. The token isn't the dollar, but holding the token activates your right to get dollars. A binding contract makes token-holding the trigger for enforceable claims against the issuer.
Legally constitutive tokens: An on-chain stock ledger is the paradigmatic case. If a corporation's governing documents designate the on-chain state as the authoritative record of holder rights, the blockchain doesn't track ownership—it constitutes ownership.
According to Shapiro, if none of these three patterns apply, it's slop. The one-line test: is the token merely a UI pointer to an off-chain registry that can ignore you when inconvenient?
The case studies Shapiro presents are overwhelming.
BitClout scraped 15,000 Twitter profiles without consent and attached bonding-curve "creator coins" to each. No enforceable right was ever granted to token holders.
Rally raised $57 million in VC funding to build a creator coin platform, then abruptly shut down its sidechain in January 2023 citing "macro headwinds." The blockchain protected nothing. The things of value—the creator's willingness to provide benefits, a functioning platform—had never existed on-chain in the first place.
The Moonbirds case is peculiar. Minted in April 2022 at 2.5 ETH, reaching a floor of 30 ETH, with holders promised commercial IP rights. Then in August of the same year, Kevin Rose unilaterally switched to a CC0 (Creative Commons Zero) license. Anyone—holder or not—could use any Moonbird image for any purpose, making the legal standing of someone who paid 30 ETH identical to someone who paid nothing. The key lesson isn't the switch itself, but that the switch was possible at all. IP rights had never been embedded in the token in any legally operative way. The issuer's ToS carried the rights, and the ToS could be changed without holder consent.
BAYC is somewhat better. Yuga Labs' terms state that "when you purchase an NFT, you own the underlying Bored Ape, the Art, completely," linking an unlimited commercial use license to cryptographic ownership verification. Holders have launched restaurants, seltzer brands, and TV shows. The concession that "at no point may we seize, freeze, or otherwise modify the ownership of any Bored Ape" is also key. It isn't perfect, but a binding legal agreement making token-holding the trigger for enforceable commercial rights puts it in a fundamentally different category from Moonbirds.
Nasdaq's September 2025 proposed rule change is also mentioned—settling equity trades in "tokenized form," except the blockchain doesn't determine price, match orders, settle trades, or hold assets. It receives a notification after the fact from the same centralized systems. The press will write "Nasdaq brings stock trading on-chain," but the reality is a notification layer for a system that functions identically without it.
BlackRock's BUIDL is a somewhat better case. There is real utility: 24/7 transfers, programmable collateral, collateral acceptance on Deribit and Binance. But the underlying assets are held by BNY Mellon, managed by BlackRock, and if Securitize disappears, the token doesn't self-execute a bond redemption.
Shapiro's argument gets genuinely scary when he points out that this problem extends beyond individual tokens to threaten blockchains themselves. Take Ethereum as an example: ETH itself is the paradigm of clean tokenization, but for ETH to be valuable, Ethereum must be useful, and for Ethereum to be useful, meaningful assets and contracts must actually settle there. If everything on the chain other than ETH is merely "the real thing lives off-chain, the token just points to it"—then what? Ethereum is just the example here, but the broader point stands: from this perspective, a blockchain becomes little more than a censorship-resistant archive of photocopies.
FLAVOR by moyedWeb Proof, Make more data verifiable
API for everything without permisson (and legally)
FLAVOR by moyed10 Weeks of Journey into vFHE
i’ve been working on deep dive into vFHE ((verifiable Fully Homomorphic Encryption)) for last 10 weeks.
FLAVOR by moyedI read Sentient Whitepaper So You don’t need to
Sentient, Platform for 'Clopen' AI Models
FLAVOR by moyedWeb Proof, Make more data verifiable
API for everything without permisson (and legally)
FLAVOR by moyed10 Weeks of Journey into vFHE
i’ve been working on deep dive into vFHE ((verifiable Fully Homomorphic Encryption)) for last 10 weeks.
FLAVOR by moyedI read Sentient Whitepaper So You don’t need to
Sentient, Platform for 'Clopen' AI Models
Support FLAVOR by moyed
FLAVOR by moyedThings I read Last Week #11
Support FLAVOR by moyed
FLAVOR by moyedThings I read Last Week #11
Articles about crypto projects that I'm interested in.
Articles about crypto projects that I'm interested in.
Subscribe to FLAVOR by moyed
Subscribe to FLAVOR by moyed
Share Dialog
Share Dialog
<100 subscribers
<100 subscribers
No activity yet