Agent discovery can use probabilistic recommendations, but credentials, payments, and checkout need deterministic flows.
2
Shared payment tokens let agents pass payment credentials with seller, amount, currency, and time limits enforced by Stripe.
3
Machine Payments Protocol and the Agentic Commerce Protocol let agents pay for APIs and complete structured commerce flows without operating a checkout page like a human.
Summary
Steve Kaliski argues that agents already spend money through model subscriptions and cloud tools, even when they do not directly pay ordinary businesses. The risk grows when an agent browses pages, extracts prices, selects credentials, and clicks through checkout. It can choose the wrong seller, product, amount, or payment method. Kaliski separates discovery, where non-deterministic language-model behavior is useful, from payment and checkout, where the steps and values need to be verifiable. Stripe's shared payment tokens apply limits to credentials and bind them to sellers. The Machine Payments Protocol lets an API request payment with an HTTP 402 response. The Agentic Commerce Protocol, built with OpenAI, gives agents structured product, cart, tax, shipping, and checkout data. The seller continues to receive payment details and risk signals. Kaliski is direct about the limits of the approach: Stripe had no public figures for adoption or payment volume yet.
Discovery can be probabilistic, while spending needs fixed rules
Kaliski separates the part of an agent that finds and recommends from the part that transacts. Language models can use broad information to recommend code, products, or businesses, so discovery benefits from non-determinism. Credentials, payments, and checkout need deterministic behavior. An agent should move from open-ended search into a programmatic flow before money is committed. He returns to this distinction at the end: discovery can remain non-deterministic, while payments, checkout, and credentials should shift toward exclusively deterministic interactions with verifiable parties and structured negotiation.
Agents already act as economic actors through the tools they use
Agents are already involved in spending when people use products such as cloud coding tools or Codex. The cost may be hidden inside a subscription or converted from model tokens into dollars, but the agent is still consuming paid resources. Kaliski's question is how to extend that capability to other businesses, currencies, payment methods, and patterns of spending. Agents produce text, read and write data, and call third-party tools. Some of those tools require payment, which makes search, credential management, and payment part of the same system.
Browser automation exposes four ways an agent can spend incorrectly
Kaliski identifies four risks in letting an agent operate a website like a human. It can buy from the wrong place, such as a fake domain that resembles Amazon. It can buy the wrong item, including one that costs much more than intended. It can spend the wrong amount because prices, taxes, currencies, or page calculations differ from what the agent extracted. It can also send the wrong credential to the wrong destination. Copying card numbers, filling forms, and clicking through pages makes these errors difficult to control and outcomes difficult to observe.
Shared payment tokens constrain a credential before a seller uses it
Shared payment tokens let an agent collect a payment credential and pass a restricted version to a seller. The token can work across hundreds of payment method types and can encode limits for a particular seller, currency, amount, and time period. In the demo, a Visa card is restricted to $25, valid for 30 days, and scoped to an internal test seller. When the seller attempts to charge $50, Stripe rejects the payment because it exceeds the mandate. After the amount is lowered, the payment succeeds. Sellers still receive details such as card brand, last four digits, and credit type for their existing risk systems.
Machine Payments Protocol connects a paid tool request to its payment
Kaliski describes the Machine Payments Protocol, developed with Tempo, for paying for HTTP-based tools. Instead of requiring an API key for every interaction, a protected endpoint can return HTTP status 402 when payment is required. The response includes an encoded payload describing what is being bought, who receives the money, and how to pay. In the demo, the request costs one cent, goes to a particular recipient in USD on the Tempo blockchain, and succeeds after approval. A transaction then appears on the blockchain, tying the payment to the requested resource.
Agentic Commerce Protocol gives checkout a structured state
For ordinary purchases, an agent needs more than permission to call an API. It must understand the product, quantity, taxes, fulfillment choices, payment method, and final amount. Stripe worked with OpenAI on the Agentic Commerce Protocol, which defines APIs and objects for these exchanges. A seller can expose a product catalog as JSON with images, descriptions, and prices. The agent selects an item and starts checkout. Each change, such as quantity, payment method, or shipping, creates a back-and-forth that returns the latest cart state. This avoids having the agent infer important details from a visual checkout page.
Businesses need agent-friendly interfaces while retaining control of the sale
Kaliski says sellers should expose structured, agent-friendly interfaces rather than only web pages or applications. Programmatic flows reduce the chance that non-deterministic browser interaction changes the purchase. Sellers still keep the customer relationship and receive payment details and risk signals needed to decide whether to accept the transaction. The system can support cards, crypto, and other payment methods. Credentials can be managed with shared payment tokens, wallets, or related mechanisms, reducing the chance that an agent spends an enormous amount on a card.
Recurring and longer-lived policies can extend the same token model
During questions, Kaliski says shared payment tokens can support recurring spending. He compares the model to giving a business a card with permission to spend a set amount periodically, while continuing to use the same underlying credential. He also compares the flow to OAuth access and refresh tokens, where later usage can be requested. For a more enduring budget, a higher limit can be set while keeping the token scoped to individual sellers. Stripe Projects is built on these shared payment token primitives, according to Kaliski.
"Discovery and exploration benefit from non-determinism, right? But credentials, payments, and checkout require determinism."00:49
Who should watch
You are building an agent that needs to call paid APIs or spend money on behalf of a person.
Your business wants agents to discover products and complete purchases without relying on browser automation.
You need to design spending limits, seller identity checks, recurring budgets, or structured checkout data for autonomous software.
