Agent Economy with receipts

Protocols · Verifiable Intent & Agent Pay

What are Verifiable Intent and Agent Pay?

Together they are the card-network payment rail for agentic commerce, backed by Mastercard. Verifiable Intent proves what the buyer authorized. Agent Pay moves the money as a card payment. Proof first, payment second.

What is it?

Verifiable Intent is a card-network specification for proving what a buyer authorized: a checkable order form the network can inspect before money moves. When an AI agent shops for a person, nobody at the checkout saw that person approve anything. Mastercard's announcement opens with the question this creates: “How do we know an agent is doing exactly what we asked — and nothing more?” Verifiable Intent answers it by turning the approval into a signed record the network can check (Mastercard announcement).

Agent Pay is Mastercard's program for agent-initiated card payments: card payments where the buyer pressing the button is authorized software. It requires agents to be registered and verified before they can pay, and it introduces Agentic Tokens — payment credentials that stand in for the real card number, built on the tokenization Mastercard already uses for contactless payments and card-on-file billing (Agent Pay announcement).

The two pair by design: Agent Pay carries the payment; Verifiable Intent carries the proof of what was approved. The specification covers the credential format, the constraint rules, and how a verifier checks the chain. It declares transport, key management, dispute resolution, and regulatory compliance out of scope (specification repository).

Who is behind it?

Mastercard, in collaboration with Google. The two introduced Verifiable Intent on 2026-03-05 as an open, standards-based way to prove authorization in agentic commerce (Mastercard announcement). It is designed to align with Google's AP2 & UCP — and to be protocol-agnostic: the specification publishes integration mappings for AP2, ACP, and UCP (verifiableintent.dev).

The specification is maintained by Mastercard and open to multi-stakeholder contribution. It builds on published standards from the FIDO Alliance, EMVCo, the IETF, and the W3C (specification repository; announcement). The March announcement carried supporting statements from payment companies including Adyen, Checkout.com, Fiserv, IBM, and Worldpay.

Agent Pay is Mastercard's own program, announced 2025-04-29 with Microsoft and IBM among the named collaborators (Agent Pay announcement).

How does a payment flow?

Verifiable Intent describes a chain of signed credentials in three layers. Each layer cryptographically constrains the next (verifiableintent.dev).

  1. The issuer signs the identity. A credential provider — typically the card issuer — signs a credential tying the cardholder to their card. This is the chain's identity layer.
  2. The person signs the terms. The cardholder delegates to the agent in a second signed credential: what may be bought, from which merchants, up to what amount. The specification defines eight constraint types, including amount bounds, merchant allowlists, and budget caps.
  3. The agent proves its scope. In immediate mode, the person approved one specific checkout — a human-in-the-loop purchase — and the credential expires within minutes. In autonomous mode, the agent acts alone inside its constraints, on a credential that lives for days or weeks (specification repository).
  4. The payment is a card payment. The agent transacts with the merchant over Mastercard's network. Under Agent Pay, it pays with an Agentic Token — a tokenized card credential issued to a registered, verified agent (Agent Pay announcement).
  5. A record remains. The result links identity, intent, and action in one tamper-resistant record. Selective disclosure keeps it private: only the minimum necessary detail is shared across parties, and only when needed (Mastercard announcement).

Because the money moves as a card payment, card-network dispute machinery still applies. Mastercard's announcement puts it plainly: “If a dispute occurs, all parties can rely on a clear audit trail.”

Where does it stand today?

The specification is young and labeled as such: Draft v0.1, open-sourced 2026-03-05 under the Apache 2.0 license — an open standard in draft form — published at verifiableintent.dev and on GitHub with a Python reference implementation and runnable examples.

The connection to Agent Pay is announced, not yet shipped as a public API. Mastercard said in March that Verifiable Intent “will be integrated directly into Mastercard Agent Pay's intent APIs,” with developer tools to follow on Mastercard Developers (announcement). As of 2026-09-02, that page still describes them as coming soon; a general-availability date is not yet published by Mastercard.

Agent Pay itself has launched in stages: the program was announced 2025-04-29 (Agent Pay announcement), and Agent Pay for Machines — extending it to machine-to-machine payments between software systems — launched 2026-06-10 (Agent Pay for Machines press release).

What this page cannot tell you is how much money moves over this rail. Mastercard's network settles privately, so there is no public settlement data for anyone to recompute. Usage figures can only come from what Mastercard publishes, and none are published yet. That is a fact about measurability, not a ranking — the bar any number must clear on this site is in the methodology.

Where are the primary sources?

Everything factual on this page traces to five documents, each from Mastercard or a Mastercard-maintained project:

The other three rails this site covers, at the same depth: AP2 & UCP, Stripe machine payments, and x402. Side by side, with every cell cited: the comparison matrix.

What changed on this page?