The scenario that payments executives have discussed in theory for two years is now running in production: software agents that hold balances, quote prices to each other, and settle transactions with no human clicking an approval button. Coverage this week describes AI agents beginning to transact autonomously, and the payments industry — Visa, Mastercard and Ant International included — is drafting standards for a future in which artificial intelligence does more than recommend products. The mechanics of how a machine actually pays, and on which rails, is the part worth examining closely, because the design choices being made now will determine who owns machine commerce.
What an Agent Payment Actually Requires
A human payment bundles several functions that an autonomous payment has to decompose. First, identity: something has to answer “who or what is paying,” which is harder for an agent than a person because agents are software that can be spawned, copied and modified. Second, authorization scope: an agent needs a wallet or credential with defined limits — a spending cap, an allowlist of counterparties, an expiry. Third, machine-readable commerce: an agent cannot browse a checkout page, so it needs a structured offer — price, terms, endpoint — it can evaluate programmatically. Stablecoin wallets with programmatic spend policies handle the second requirement natively; the first and third are where emerging protocols and the card networks’ standards work overlap.
The Stablecoin-Native Stack
The agent-native approach treats the wallet as the agent’s bank account. Coinbase’s Agent Payments toolkit gives developers the pieces to let an agent hold and spend USDC, and x402 — an open payment protocol built around HTTP — lets a server state a price for a resource and an agent pay it programmatically, with settlement in stablecoins. Skyfire and Payman sit in a similar space, providing payment infrastructure and spend controls for autonomous agents. The appeal is straightforward: stablecoins settle in seconds, cost fractions of a cent regardless of amount, work in amounts too small for card rails to tolerate, and require no acquiring bank to onboard a machine counterparty. The sub-cent transactions that dominate current machine-to-machine volume are only economical on rails priced for them.
The Card-Rail Counterargument
The card networks are not conceding the identity layer. As we covered earlier this week, Visa, Mastercard and Ant International are co-developing standards for identifying and monitoring agents that initiate payments, and Mastercard’s Agent Connect already wires merchants into a directory that shopping agents can browse. The card pitch is that merchants, issuers and fraud systems already exist on those rails, and consumer-facing agent purchases — an AI assistant booking a flight — will route through credentials humans already trust. The tradeoff is cost and granularity: interchange economics and chargeback machinery are a poor fit for high-frequency, low-value machine transactions. The likely split is consumer-facing agent purchases on cards, machine-to-machine settlement on stablecoins.
Open Questions
Three problems remain unresolved. Agent identity has no standard: the card networks’ approach and agent-native protocols do not interoperate yet, and a fragmented identity layer raises fraud costs for everyone. Liability is undefined in most jurisdictions — when an agent exceeds its mandate and mispays, existing consumer-protection and contract law does not cleanly assign responsibility, a gap regulators outside the U.S. have started to flag. And spend-control tooling is immature: policy enforcement at the wallet level is exactly the capability that determines whether a capped, sandboxed agent is safe to give money to. Protocol design here is doing the work regulation has not yet done.