Now the agents pay each other.

Most of the agentic-commerce story so far has been about a person at one end and a shop at the other, with an agent doing the clicking in between. This week the picture changed. The newest move is not an agent buying from a merchant, it is an agent paying another agent, at machine speed, in amounts a card network was never built to carry. Here is what landed, and what it means from the rails.

Machine pays machine

Mastercard launched Agent Pay for Machines on 10 June, a payment framework aimed at a world where software buys services from other software [1]. The pitch is sums as small as fractions of a cent, completed programmatically and in chains, so an agent can pay for an API call, a data lookup, or a compute slice as it works [2]. More than thirty partners signed on at launch, spanning processors such as Stripe, Adyen, and Checkout.com and crypto firms such as Coinbase, and the protocol is built to settle across cards, bank rails, and stablecoins rather than one of them [1].

The consumer side kept filling in

The human-facing rails moved too. Visa tied its agentic push to OpenAI so that tokenised, agent-led purchases can happen inside the model itself [3]. Adyen unveiled a suite called Adyen Agentic on 16 June, covering the stages of an agent-driven customer journey for enterprise merchants [4]. Demand is not hypothetical. Visa's research has 47% of US shoppers already using AI for at least one shopping task, such as price comparison or recommendations [4].

Someone finally named the hard part

At NY Tech Week, JPMorgan's payments team put the uncomfortable question on the table: the real problem is not whether a machine can pay, it is deciding what you are willing to delegate to one, at what threshold, and under what conditions [5]. Their other point matters more than it sounds. For any high-value payment, you need auditability, the ability to say after the fact why the payment was made and under what policy, and they framed that as the foundation an agentic system has to sit on, not a feature bolted on later [5].

And the money underneath kept getting a rulebook

While the front end learned to spend, the back end stayed in active rulemaking. The GENIUS Act, signed into law on 18 July 2025, is now being written into regulation. Treasury's proposal on how it will judge whether a state regime is substantially similar to the federal one closed for comment on 2 June, and the joint FinCEN and OFAC rule treating permitted stablecoin issuers as financial institutions under the Bank Secrecy Act closed on 9 June [6]. Final regulations are due by 18 July 2026 [7].

Read from the rails

Machine-to-machine micropayments are where the old assumptions break first. A card transaction assumes one buyer, one merchant, one authorisation a human could in principle inspect. An agent paying thousands of fractional charges an hour assumes none of that, and it moves the whole system from "approve this purchase" to "did the policy hold across the run." That is closer to how a payments operation already thinks than how a checkout button thinks.

The parts I would watch are the parts the launch slides skip. Replay protection: an agent that retries a failed call must not pay twice, which is the same idempotency problem I have hit running my own agents, and at fractions of a cent per call it hides easily until the totals do not reconcile. Auditability: JPMorgan is right that you need to prove why a payment fired and under what mandate, and that proof has to survive into the dispute, not just the dashboard. And exception handling, because a protocol is judged not on the day it launches but on the first month of edge cases. The networks have built the happy path. The repair queue is where this gets won.

Agent-to-agent payments turn the question from "approve this purchase" into "did the policy hold across the run." That is an operations problem, and operations is judged on the exceptions.