A Federal Reserve governor stood up last week and described a payments problem without ever using the word "payments." He called it a market structure issue. That vocabulary choice is the most tradeable signal in the entire discourse.
Visa had already published a technical specification under the Intelligent Commerce banner. Mastercard had extended Agent Pay with probabilistic scoring and a new class of trust signals. Neither announcement moved a token. Both quietly shipped the artifact that actually matters: a machine-readable delegation record. A structured object that names the principal, the spending limit, the merchant scope, and the expiry. Read that object carefully and you can see where the next decade of payment margin gets collected. Follow the gas, not the hype.
I have spent enough time inside settlement logic to know that the boring schema always outlives the exciting narrative. In 2019 I reverse-engineered Uniswap v2 pricing and found an oracle edge case that only mattered under volatility. The finding was small. The update it forced was permanent. Agentic payments are at that same stage. The schema is being written now, and almost nobody is reading it.
Here is the essential background, stripped of the press release. Two card networks — Visa and Mastercard — are competing to become the trust layer for transactions initiated by autonomous software agents rather than humans. The premise is simple. A user grants an AI agent a bounded mandate. The agent then purchases on the user's behalf: SaaS subscriptions, API calls, compute, supply-chain replenishment, price comparisons. The agent needs a rail. The card networks want to be that rail.
The Fed's framing matters because it defines the regulatory envelope. By calling this a market structure question rather than an infrastructure supervision question, the central bank is deliberately stepping back. It is drawing a map and pointing the industry toward self-governance — certification regimes, liability rules, and trust signals developed by the networks themselves.
The technical pivot is equally clear. Traditional card rails were built for human cadence: a few transactions per day, each authenticated by a person. Agentic commerce inverts that. Machine-to-machine payments run at high frequency, small value, and unpredictable timing. The economics of a $0.003 API call cannot survive a fixed interchange structure designed for a $40 dinner. Machine-to-machine traffic prefers low fixed-fee rails. That single constraint reshapes clearing, settlement, and fraud.
So the networks are doing two things at once. They are extending their authorization standard to accept machine-readable mandates. And they are inventing a scoring layer — probabilistic trust — that converts a binary approve/decline into a weighted settlement decision. The clearing logic is migrating from "did it succeed" to "how much do we trust the agent that asked."
The early penetration is B2B, not consumer. Corporate procurement has natural guardrails — budget ceilings, approved vendor lists, recurring orders — that make a machine mandate legible to an issuer. A consumer handing an agent their grocery budget is a trust problem with no audit trail. A procurement manager handing an agent a replenishment mandate is a policy object. The networks know this, which is why the first real volume will look like enterprise software, not a wallet.
Let me go where the analysis usually stops: the actual mechanics. Start with the mandate. A machine-readable delegation record is not a contract in the legal sense. It is a data structure. It contains a principal identifier, a scope of permitted merchants, a ceiling, a time window, and — critically — a revocation field. When an agent presents this object, the issuer does not evaluate intent. It evaluates conformance. Does the request fall inside the box the human drew? Yes or no.
This is elegant, and it is also the first crack. The box is only as good as the human who drew it. A principal who sets a $500 ceiling and an "any merchant" scope has effectively delegated unbounded discretion up to that ceiling. Now multiply that across ten thousand mandates. You have created a fleet of semi-autonomous spenders, each individually rational, collectively opaque.
Mastercard's answer is probabilistic scoring plus trust signals. Instead of a binary decision, the network attaches a confidence value. A high-confidence agent settles at a favorable rate. A low-confidence agent triggers step-up verification or rejection. On paper this is a risk model. In practice it is a pricing mechanism. The trust score becomes the new interchange tier. Whoever owns the scoring function owns the margin. Alpha hides in the margins, and the margin here is a probability.
Now watch the settlement side. High-frequency, low-value machine payments push clearing toward real-time gross settlement. That is not a feature. It is a liquidity requirement. In a T+1 world, a bank can net flows across a day and fund the residual. In a millisecond world, there is no netting window. Every micro-payment is its own settlement event. Liquidity management shifts from daily to continuous, and the reserve buffer that banks must hold against agentic traffic rises sharply.
I have seen this pattern before. When I built a Python scraper across Compound and Aave in 2020, the arbitrage window on sETH yields lasted exactly seventy-two hours. It closed not because the opportunity vanished, but because enough capital arrived to compress it to zero. Agentic payment rails will follow the same curve. The early fee advantage of a low-fixed-cost machine rail will attract volume until the spread collapses. The durable value is not the rail. It is the trust score that gates it.
Then there is fraud. Governor Waller was explicit: fraud and monitoring systems must be recalibrated for autonomous agents. He is right, and the reason is structural. Traditional anti-money-laundering logic profiles human behavior — velocity, geography, counterparty patterns. An AI agent has no geography. Its velocity is programmable. Its counterparty set is whatever the mandate permits. The behavioral fingerprint that classical AML depends on simply does not exist for a machine principal.
Worse, the attack surface is new. Prompt injection — manipulating an agent's instructions through crafted inputs — can convert a compliant spender into a fraudulent one without ever touching credentials. The mandate stays valid. The agent's interpretation of it does not. This is operational risk migrating from the insider to the adversarial model, and the existing business-continuity frameworks were never built to price it.
Here is the part the market is not pricing. The liability question is unanswered, and the card networks know it. When an agent buys the wrong thing — a hallucinated bulk order, a misread quantity — who eats the loss? The principal claims overreach. The agent vendor claims the mandate authorized it. The issuer is holding a settlement obligation to a merchant who delivered exactly what was ordered. Someone must absorb the loss first, and in practice that someone is the issuer, because the settlement clock does not pause for a liability dispute. That is a liquidity mismatch dressed as a legal question.
There is a quieter problem buried in the mandate. Agentic decisions require behavioral context — purchase history, preferences, timing. The more context the agent carries, the better its decisions and the larger the data surface it exposes. High-frequency machine payments generate a data stream that no human review process can audit at the same clock speed. Consent frameworks built for a click do not translate to a decision made in four hundred milliseconds. The mandate authorizes the spend. It does not, by itself, authorize the inference.
And watch the concentration. If a handful of model providers supply the decision layer, the networks become dependent on their interface standards. A single dominant model that defines how mandates are expressed can effectively set the network's schema. That is technical lock-in disguised as interoperability. The scoring monopoly and the model monopoly can merge, and the entity that controls both controls the toll.
The consensus narrative is that card networks are the natural winners here. They have the network, the rules, and the merchant acceptance. I am less certain, and the reason is vertical integration. The value in agentic commerce sits at two chokepoints: the mandate and the intelligence. The networks are fighting hard for the mandate. But the intelligence — the model that decides what to buy — belongs to a handful of AI firms. If those firms build their own payment protocol and connect directly to banks, the card network becomes a pipe. A pipe earns throughput fees. It does not earn trust rents.
The open-versus-closed standard fight is therefore not academic. An open standard preserves the networks as neutral infrastructure. A closed ecosystem — a walled agentic wallet tied to one model provider — routes around them. The working groups forming now will decide which world arrives. And correlation is not causation: a network publishing a specification is not the same as a network controlling the standard. Publishing is cheap. Adoption is not.
B2C is where the narrative lives, and B2C is where the liability is thinnest. A consumer agent that misreads a request produces a dispute a human can escalate. A machine fleet misreading the same request produces a correlated loss that no dispute process can absorb. The consumer story sells the vision. The enterprise story pays the bills. Do not confuse the two when you underwrite the rails.
In a bear market this matters more, not less. Infra narratives get repriced when capital is scarce. The projects that survive are the ones whose fee capture is structural rather than promotional. Agentic payments are one of the few rails where the fee is attached to a non-discretionary need — software has to pay for compute whether the market is up or down. That is the kind of demand that does not need a bull market to justify its existence.
Watch one signal over the next two quarters. If the AI model providers join the open standard and interoperate, the networks keep their position and agentic payments become a volume story. If a major model provider announces a proprietary settlement path, the networks become utilities and their multiple compresses. Code does not lie; people do. The schema will tell you which world you are in before the price does.