On September 10, 2026, three logos shared a stage in São Paulo — Visa, Mastercard, Ant International — and between them announced something called KYA, Know Your Agent, a framework that gives autonomous AI agents a verifiable identity capable of traveling across payment networks. No regulator was in the room. No central bank observer. No politician. The silence was the first data point of the morning, and it read louder than any floor price on the screens behind them.
I have spent twenty-three years learning to distrust the number quoted on launch day and to hunt instead for the number buried in the footnote. The launch quoted McKinsey: three to five trillion dollars in agentic commerce by 2030. The footnote said fourteen percent. That is the share of consumers who claim they would trust an AI to complete a purchase without any verification layer. Somewhere between the trillions and the fourteen percent sits the entire unresolved story of KYA, and I want to walk through the architecture before the narrative sets.
What KYA actually is
KYA is not a product. It is a translation layer. Underneath it sit three proprietary protocols that, until that September morning, had no reason to speak to one another. Visa's TAP ships with twelve partners, among them Adyen, Shopify, and Stripe. Mastercard's Verifiable Intent, open-sourced and co-built with Google, answers a harder question than identity — not which agent this is, but whether the human actually meant this. Ant International brings AMP and something neither American network can replicate alone: a wallet ecosystem touching roughly thirteen trillion dollars in spend, concentrated in the emerging markets where card penetration never fully arrived.
The stated promise is interoperability. Register an agent's identity once and it should be recognized across all three networks without re-onboarding. That promise is genuinely valuable at the developer and merchant margin — it collapses the marginal cost of connecting an agent to commerce the way a shared language collapses the cost of translation. But register once is doing quiet, load-bearing work in that sentence. Behind it must stand a registry, either a single source of truth or a federation of them, that issues and revokes agent credentials. The announcement did not say who runs it.
I have watched this shape before. In 2017, auditing an ICO's token distribution contracts in Chengdu, I found an integer overflow that could have drained fifteen percent of the raise. The team wanted to ship; I insisted on a patch and cost them three days. The lesson that stayed with me was never about arithmetic. It was that the most dangerous component of any system is usually the one everyone has silently agreed not to discuss yet. For KYA, that component is the root of trust.
The registry is the real product
Tracing the ghost in this architecture leads to a single question: who holds the issuance authority? W3C Verifiable Credentials and DID-style identifiers almost certainly form the substrate, because the consent-based, minimal-disclosure language in Mastercard's materials points there. Protocol translation, hard as it sounds, is a solvable engineering problem. Issuance authority is a governance problem, and governance problems do not compile.
Whoever operates the unified registry inherits something that behaves like a certificate authority — the power to decide which agents exist in the eyes of the commercial internet, and the power to charge rent on that existence. The three participants have deliberately left this ambiguity in place, which tells me the ownership has not been settled. This is not an oversight. It is the unresolved bargaining chip that every future negotiation between the three will orbit. Silence speaks louder than floor prices, and this particular silence is worth billions.
The numbers hold the memory that the press release ignores. Consider the twenty-five-dollar threshold. Forty-two percent of consumers say they will not trust an AI purchase above that figure. That number is not a psychology artifact; it is chargeback economics wearing a survey result. Below twenty-five dollars, the cost of disputing a transaction exceeds the value at stake, so the loss can be absorbed quietly. Above it, disputes multiply and — critically — nobody has defined who eats them. The credit risk of agent commerce is not a creditworthiness problem. It is a liability-allocation problem, and the framework's own interoperability makes it worse, because responsibility can now be passed among three networks instead of resting on one.
Mapping the invisible currents of that liability is where the framework is thinnest. Agent payments are transactional and immediate, so classic liquidity risk barely applies. But give an agent the ability to hold balances or pre-authorize recurring spend — a small product extension, not a redesign — and you have manufactured a new pool of floating value with no defined custodian. The framework does not touch this today. It does not have to. The next product cycle will.
The blast radius problem
Here is the architectural cost that interoperability hides. Today, a compromised credential damages one network. Under KYA, a hijacked agent carries a legitimate identity across three networks simultaneously, and no synchronized freeze mechanism exists to revoke it everywhere at once. I spent 2022 mapping the on-chain drain of TerraUSD across half a million micro-transactions in the forty-eight hours before it broke, and the pattern I remember is not the price. It is how fast a coordinated system becomes a coordinated liability when the emergency brake was never specified. Interoperability converts a single-point breach into a three-network event. That is the trade being made: efficiency purchased with resilience.
The same gap opens in compliance. KYA extends know-your-customer duties from people to non-human agents — registration, authentication, continuous monitoring. That is a genuine anti-money-laundering upgrade. But when a hijacked agent moves a suspicious transfer, who owes the report? The originating network, the receiving network, or the registry that minted the identity? The framework's very interoperability creates space for each party to assume another is handling it. That space is a regulatory arbitrage window, and it will not stay open forever.
Then there is the biometric anchor. Visa's two-point-four-billion-dollar acquisition of BioCatch is the tell. Behavioral biometrics are classified as sensitive personal information under both GDPR and China's PIPL, demanding separate consent and a higher bar. Embedding them into the agent identity chain means the trust that authorizes an autonomous machine ultimately rests on the measured behavior of a human body. That raises the consequence of any breach from a token leaked to a person's physical signature leaked. The privacy cost of KYA is being quietly shifted from the network to the individual.
Coloring the grey area of who actually competes
The competitive map most observers drew is wrong. The three networks are not primarily competing with one another. They aligned precisely because they share an external threat: the agent platforms themselves. If OpenAI or Google hosts identity and payment natively inside the agent runtime, the card network's identity layer becomes a backend query that may never be reached — a component that exists but is bypassed. Visa and Mastercard inviting Google into Verifiable Intent is a defensive embrace, turning a potential disruptor into a partner. It is also, by design, letting the fox inspect the henhouse.
Within that alignment, Ant International occupies the most exposed seat. It commands the strongest distribution — the wallet ecosystem is a moat neither American network can buy — yet it holds the weakest voice in the cryptographic standards layer where identity and intent are codified. Being the network that translates protocols, rather than the one that authors them, risks a slow demotion: from an alternative rail into a subordinate node of the card networks. Watching the block confirm, not the narrative, I would want to see Ant's name on the registry governance documents before I believed the partnership is symmetrical.
The contrarian read
Correlation is not causation, and a joint press conference is not a market. Three networks that could have spent a decade bleeding each other with rival standards instead chose alignment. That is not the natural drift of competition. It is oligopolistic pre-positioning against a shared threat, dressed in the language of open infrastructure. When incumbents describe interoperability, read it as: we will define the interoperability.
The second uncomfortable read concerns the network effect itself. We are told more participants make the system more valuable. True — until the three principals lock the registry, at which point the network effect stops being an open dividend and becomes an exclusionary wall. The critical mass, once reached, protects the founders and evicts the latecomers: regional wallets, independent identity vendors, the next Ant.
And the entire value case rests on a forecast rather than a demand signal. Three to five trillion by 2030 is a prediction, not a receipt. AI-driven retail traffic jumped forty-seven hundred percent, but from a base so small that the growth describes experimentation, not adoption. Supply is pulling this market. Consumers are not yet pushing it. When the ratio between those two forces stays that wide, you are not looking at a market. You are looking at a rehearsal.
The pattern emerges in the quiet hours
What I will watch over the next two quarters is not another partnership announcement. I will watch where the registry lands — which entity files the governance charter, and whether Ant signs as an author or a tenant. I will watch Brazil, because São Paulo was chosen over New York and Beijing for a reason: emerging markets, where wallets leapfrogged cards and regulators are still writing the rules, are the lowest-friction place to set a global default. And I will watch the first time a hijacked agent moves funds across three networks in a single hour, because that is the event that will force the responsibility question out of the appendix and onto the front page.
The framework will likely survive. Whether the people who use it are served, or merely enrolled, depends on a document no one has published yet.