Building safe Payment Infrastructure for the autonomous economy — Steve Kaliski, Stripe

TL;DR

• 1

  LLMs can explore non-deterministically, but payments cannot: Kaliski's core claim is that discovery benefits from agent flexibility, while credentials, checkout, and money movement need deterministic, verifiable flows.

• 2

  Shared payment tokens let robots spend without exposing raw cards: Stripe can issue a token scoped to a specific seller, amount, currency, and time window, so a Visa with a high credit limit can be safely constrained to something like $25 for 30 days.

• 3

  Stripe enforces the policy even if the seller asks for more: In the demo, a seller tries to charge $50 against a token capped at $25, and the payment fails because the limit is enforced by Stripe, not by trust.

• 4

  Paid tool calls should look like native internet payments: Working with Tempo, Stripe uses a machine payments protocol where an HTTP 402 response tells an agent what resource costs, who gets paid, and how to pay before retrying the request.

• 5

  Agent checkout needs structured commerce data, not browser automation: With the agent-to-commerce protocol built with OpenAI, sellers expose catalogs, cart state, tax, shipping, and payment options as JSON so agents do not have to scrape websites and guess.

• 6

  Stripe is already treating agents as economic actors: Kaliski notes that agents already "spend money" through tokens and subscriptions to model providers, and the real shift is extending that spending power to the rest of the economy safely.