Skip to content

From October 1, WhatsApp charges for every business reply. Read the migration playbook

Who pays when the agent buys? AP2, x402, ACP and MPP, plainly

Four protocols, four layers of the same problem: letting software spend money on your behalf without it becoming a disaster. Here is what each is actually for, and what is still genuinely unsettled.

Who pays when the agent buys? AP2, x402, ACP and MPP, plainly

Agentic commerce is the idea that an agent takes a goal rather than a click and runs the whole purchase — discovery, authorization, payment, fulfilment. The interesting engineering is not the buying. It is proving that the human meant it, and deciding who is liable when the machine got it wrong.

The four names you will keep hearing

  • AP2, the Agent Payments Protocol — introduced by Google Cloud and Coinbase in September 2025 with a large partner list. It is the authorization and trust layer: cryptographically signed mandates that record what the human actually authorised. Supports card rails, and crypto through an x402 extension.
  • x402 — from Coinbase, reviving the long-dormant HTTP 402 status code for stablecoin payments over HTTP. Built for machine-to-machine microtransactions and API monetization. It has the most production traction of the four: V2 landed in December 2025, Stripe integrated it in February 2026, and Cloudflare supports it.
  • ACP — the checkout layer for conventional e-commerce flows.
  • MPP — machine-to-machine payments, in the same neighbourhood as x402.

They are layers, not rivals

The framing that trips people up is treating this as a format war. A real system might use AP2 to capture and verify the authorization, ACP to complete a retail checkout, and x402 or MPP for the machine-to-machine calls underneath. Asking which one wins is like asking whether OAuth beat HTTP.

What is genuinely unsettled

Security research on x402 is already being published, including systematic analyses of attacks against it — which is healthy and also a signal of maturity. Signed mandates raise a harder question than the cryptography: what does a signed intent mean when the model misread the goal it was given? And nobody has a clean answer for disputes and refunds when the counterparty was a machine acting under delegated authority.

Where we stand

Fiat first. An agent that can raise a payment request inside a conversation and have it reconciled through the platform's existing billing covers the overwhelming majority of what our customers actually need today — a consultant taking a deposit, a shop closing a sale in chat. We are watching the protocols closely and we are not asking anyone to bet their checkout on one of them yet.

Keep reading

Related guides

Industry news

MCP is not one vendor's protocol any more

Anthropic donated the Model Context Protocol to the Agentic AI Foundation under the Linux Foundation in December 2025. If lock-in was your objection to building on MCP, it no longer holds.

Weighing up your options?

Customer support

Eight tools people actually move to when their support AI stops fitting — what each is genuinely good at, what it costs you in flexibility, and how to pick without a three-month evaluation.

Sales

Eight tools for putting an AI in front of a buyer — qualifying, answering the awkward questions and getting a named contact into your pipeline rather than an anonymous chat log.

Human ambition. SuperCognit intelligence.

Try what you just read about

Free for 7 days — bring a prompt, leave with a product on your own domain.