E-commerce assumed a human at a keyboard. Every fraud model, checkout flow and chargeback rule derives from that assumption. Agents break it: a program can visit a hundred stores in a minute, compare prices, and complete purchases — and the merchant's fraud systems see traffic that looks like a bot but carries legitimate payment credentials.
The industry's answer is delegation infrastructure. AP2 (Agent Payments Protocol, developed by Google with a broad coalition including payments companies and major card networks) defines cryptographically signed 'mandates' that capture what the user authorized, when, and with what constraints. x402 (originated at Coinbase) takes a different angle: an HTTP-native payment layer enabling machine-to-machine micropayments, priced for agents that pay per request rather than per subscription.
Why it matters
Commerce is the largest economic surface AI touches, and the agent channel changes who holds the customer relationship. If a shopping agent mediates the transaction, the brand's site becomes an API — and discovery, pricing and loyalty all shift toward whichever agent the user trusts.
For merchants, the near-term issue is fraud economics. Agent traffic is indistinguishable from bot traffic without protocol support, and liability rules for agent-initiated purchases (who bears a fraudulent mandate? the user, the agent provider, the issuer?) are being written now, in protocol specifications and network rules rather than in legislation.
How the pieces fit
The stack has layers: identity (which agent, acting for which human), intent (what the user actually asked for, captured in a signed mandate), product data (machine-readable catalogs that agents can parse reliably), and settlement (the actual payment, which may be a card credential, a stablecoin transfer, or a bank rail). Protocols differ mainly in how much of this they standardize.
Card networks are participating rather than resisting: their programs for agentic commerce define how agent-initiated transactions flow through existing authorization systems, with tokens distinguishing delegated purchases from ordinary card-not-present transactions. The incumbent rails adapt; they do not disappear.
Evidence
Protocol adoption is measurable: AP2 has published extensions for multiple payment methods and lists partners across card networks, payment processors and major merchants; x402 implementations power machine-to-machine payments in crypto-native contexts. Anthropic and OpenAI have both shipped or announced commerce features for their assistants, creating demand for exactly this plumbing.
Transaction volumes through agent channels remain small relative to total e-commerce, but the direction is unambiguous — every major payments player has an agent commerce strategy, which is itself the strongest signal that the infrastructure layer is becoming real.
The competing read
Skeptics note that consumers have been promised frictionless automated shopping before, and that agent commerce may concentrate in narrow use cases — replenishment, routine purchases — while high-consideration purchases stay human-mediated. If agents only handle the boring transactions, the protocol war matters less than projected.
Protocol partisans counter that the boring transactions are enormous — subscriptions, replenishment, B2B procurement, travel bookings — and that agent-mediated discovery will reshape marketing economics even where purchases stay human-touched. Both readings agree that standards matter: fragmentation across incompatible protocols is the scenario everyone loses.
What happens next
Watch merchant adoption of the catalog and mandate standards, watch card-network rules for agent-initiated transactions (liability allocation is the pivotal detail), and watch whether a dominant agent interface emerges — the protocol is plumbing, but the consumer-facing agent is the brand. Payments infrastructure is historically slow to change and nearly impossible to change back; the specifications being written this year will outlive today's model landscape.