World and Coinbase launched a toolkit that lets AI agents carry cryptographic proof they are backed by a verified human. Nearly eighteen million people have scanned their irises to create that proof. The identity layer for agent commerce now exists. The authorization layer still does not.
On March 17, World — the identity company co-founded by Sam Altman — launched AgentKit in partnership with Coinbase. The toolkit lets AI agents carry cryptographic proof that they are backed by a unique, verified human. A developer building a shopping agent, a booking agent, or a financial agent can now integrate World ID credentials so that the agent presents proof-of-personhood to any platform it interacts with. No personal information is transmitted. The proof is zero-knowledge: the platform learns that a real human is behind this agent, and nothing else.
Tiago Sada, Chief Product Officer at Tools for Humanity — World's parent company — compared the functionality to delegating power of attorney. A verified human delegates identity credentials to an AI agent. The agent acts on the human's behalf. The delegation is cryptographic, not behavioral.
The timing is not accidental. Two days ago, The Attestation documented the first specification designed to prove a human authorized a specific transaction. Today, the industry's response is a layer beneath that: proving there is a human at all.
The Iris
World's verification method is biometric. Users visit a physical device called an Orb — a proprietary iris scanner deployed across a hundred sixty countries. The Orb captures the geometry of the iris and converts it into an encrypted numerical string stored on the user's device. World claims it does not retain the biometric data. The result is a verified World ID: a credential that proves the holder is a unique human being without revealing which one.
Nearly seventeen point nine million people have completed this process. Each receives a World ID that can be delegated to multiple AI agents through AgentKit's credential framework. The delegation uses zero-knowledge proofs — cryptographic constructions that allow one party to prove a statement is true without revealing any information beyond the truth of that statement. An agent carrying a delegated World ID can prove to a merchant that it represents a verified human without disclosing name, location, age, or any other attribute.
This solves a real problem. As agent traffic scales, platforms face a flood of automated requests they cannot distinguish from legitimate human activity. Ticket scalpers, reservation hoarders, and bot swarms look identical to an AI shopping agent acting on behalf of a real person. AgentKit gives platforms a filter: agents backed by verified humans get through. Agents without that proof do not.
The filter is bound per human, not per agent. One verified World ID can delegate to multiple agents, but a platform can cap total activity per underlying human — preventing a single person from circumventing restrictions by spawning thousands of agents. The binding runs deeper than a login. It is a cryptographic tether from silicon to iris.
The Pipe
AgentKit ships with an integration into Coinbase's x402 protocol — the payment infrastructure that revives the long-dormant HTTP 402 "Payment Required" status code as a native transaction layer for the web. When an agent requests a resource protected by x402, the server responds with a payment instruction. The agent sends a stablecoin payment — USDC on Base, Solana, or other supported chains — and receives the resource. The entire cycle runs programmatically, without a human touching a checkout form.
Erik Reppel, Head of Engineering at Coinbase's Developer Platform and founder of x402, described the integration directly: "Payments are the 'how' of agentic commerce, but identity is the 'who.' By integrating World ID with x402, developers have a complete trust stack."
The stated numbers are large. Coinbase reports over one hundred million payments processed through x402 across APIs, applications, and AI agents since launch. But the genuine daily volume tells a different story. An analysis by Artemis found that roughly half of observed x402 transactions are gamified — self-dealing, wash trading, and incentive farming rather than real commerce. Genuine daily volume runs approximately twenty-eight thousand dollars across roughly one hundred thirty-one thousand transactions, with an average payment of twenty cents.
One day in February logged three point eight million transactions and approximately two million dollars in volume — a spike that illustrates the distance between peak capacity and daily reality. The infrastructure exists before the demand. x402 is plumbing laid down before the building is constructed, a pattern common to every payment rail that eventually mattered.
The Scale Already Here
The demand, however, is arriving from directions the Western stack was not built to see.
Tearline — an AI agent infrastructure platform operating across BNB, SUI, and TON blockchains — reported nineteen point four million on-chain agent transactions with over twenty million dollars in executed task value and a ninety-six point four percent success rate. These are not test transactions. They are automated workflows, protocol interactions, and cross-platform operations executed by AI agents at production scale.
The larger signal is in China. Alipay's AI Pay — launched as part of Ant Group's Spring Festival campaign — hit one hundred twenty million transactions in a single week in February 2026, with over one hundred million users. That is three orders of magnitude beyond x402's genuine daily volume. Agentic commerce is not theoretical in China. It is operational.
McKinsey projects agentic commerce could reach three to five trillion dollars globally by 2030. Bain estimates AI agents could represent up to twenty-five percent of U.S. e-commerce by the end of the decade. During Cyber Week 2025, AI-driven interactions influenced approximately sixty-seven billion dollars in global online sales — roughly twenty percent of total digital orders. The transactions are real. The question is whether the infrastructure can make them trustworthy.
The Stack
The Payment Rail documented six protocols that launched in five months to let AI agents buy things. The Attestation documented the first specification designed to prove the human authorized the specific transaction. AgentKit adds the layer beneath both: proving the human exists.
The stack is now visible. Google's Universal Commerce Protocol — co-developed with Shopify, Target, Walmart, and Wayfair — handles how agents discover and negotiate with merchants. Stripe's integration with x402, launched February 11, handles USDC agent payments on Base. Visa's Trusted Agent Protocol verifies agent identity with over a hundred partners. Mastercard's Agent Pay, expanding across Latin America in 2026, orchestrates the payment itself. Cloudflare's x402 Foundation, co-established with Coinbase, provides the edge infrastructure.
And now World provides the identity floor: cryptographic proof that a specific, unique human is behind the agent making the request.
Each layer answers a distinct question. The commerce protocol answers: what is this agent trying to buy? The payment rail answers: how does the money move? The agent identity layer answers: is this agent legitimate? The human identity layer — AgentKit — answers: is there a real person behind it?
No single protocol answers all four questions. But the layered architecture is converging toward a system where a transaction can carry simultaneous proof of commerce intent, payment capability, agent legitimacy, and human backing. Reppel's phrase — a "complete trust stack" — describes the aspiration.
The Gap
The aspiration has a structural hole.
AgentKit proves that a verified human delegated identity to an agent. It does not prove that the human authorized the specific action the agent is taking. The delegation is general — a power of attorney, in Sada's metaphor. A power of attorney does not specify which contract to sign. It authorizes the agent to act within a domain. The specific acts within that domain are the agent's decisions, not the human's.
This is the distinction The Attestation documented in Mastercard's Verifiable Intent specification. Verifiable Intent binds a human's authorization to a specific transaction through a cryptographic delegation chain — amount bounds, merchant allowlists, budget caps, recurrence terms. Each constraint travels with the credential as a mathematical fact, not a behavioral assumption. The authorization is scoped, bounded, and auditable.
AgentKit binds a human to an agent. Verifiable Intent binds a human to an action. These are orthogonal operations.
A merchant receiving an AgentKit credential knows that the agent represents a real person. It does not know whether that person approved this purchase, at this price, from this seller. A merchant receiving a Verifiable Intent credential knows exactly what the human authorized and can verify any deviation cryptographically. The first proves identity. The second proves intent.
The gap is not theoretical. Arch Public CEO Tillman Holloway voiced the concern directly when discussing AI agents with financial autonomy: "You don't want an AI agent going, 'This is an opportunity of a lifetime — bet the farm.'" An agent carrying proof-of-human can still act outside the human's wishes. The iris scan proves who is behind the agent. It says nothing about what the agent was told to do.
World's token — WLD — has lost seventy-six percent of its value since its 2023 launch. The Register noted that neither World nor Tools for Humanity provided comment for their coverage. The crypto adoption side of World has stalled. But the identity infrastructure is real, the verification network spans a hundred sixty countries, and the problem it solves — distinguishing human-backed agents from autonomous bots — is genuine and growing.
The binding between human and agent is the first step. The binding between human and intent is the second. Today the industry shipped the first. The second remains the open layer in an assembling stack — the difference between knowing who is driving and knowing where they agreed to go.
The Assembly
Alex Blania, CEO of Tools for Humanity, framed the launch in terms of inevitability: "Proof of human is necessary to boost confidence among people ahead of a global transition towards an AI-driven era."
He is right about the necessity and incomplete about the sufficiency. Proof of human is the floor. It is not the ceiling. The agent commerce stack needs both floors and ceilings — identity that establishes who is present and authorization that constrains what they can do.
What is assembling, across six months and a dozen announcements from the largest technology and financial companies in the world, is the infrastructure for a new kind of transaction. Not human-to-merchant. Not bot-to-API. Something in between — an agent backed by a human, verified by cryptography, funded by stablecoins, authorized by a delegation chain, and accountable through an audit trail.
The Payment Rail showed the pipes. The Attestation showed the specification. The Binding shows the identity floor. Each layer that ships makes the missing layers more visible. The stack assembles from both ends — payments from below, governance from above — and the authorization layer in the middle remains the part no one has fully built.
Nineteen point four million agent transactions on three blockchains. One hundred twenty million agent payments in a single week in China. Nearly eighteen million irises scanned. The agents are transacting. The humans are being verified. The question of whether each transaction carries the human's specific intent — not just their identity, but their actual, bounded, cryptographically verifiable approval — is the question the next layer must answer.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)