Steve Kaliski from Stripe discusses building safe payment infrastructure tailored for the autonomous economy, where robots act as economic agents capable of spending money and interacting with businesses. He emphasizes the critical distinction between discovery, which benefits from non-determinism (such as AI-driven recommendations), and payments or checkout processes, which require strict determinism to ensure security and accuracy. Kaliski highlights that agents already function as economic actors, often spending money indirectly through subscriptions or tokens, and the challenge lies in enabling diverse payment methods and spend patterns safely.
Kaliski outlines the main risks when agents handle payments: purchasing from the wrong place, buying incorrect or overpriced items, spending incorrect amounts, or using wrong credentials. Traditional approaches like letting robots operate browsers mimic human behavior but are slow, error-prone, and risky. Instead, Stripe advocates for API-driven, programmatic payment flows with verifiable identities and spend policies to minimize risk and improve reliability. This approach separates the discovery phase from the transaction phase, ensuring that payments are secure and deterministic.
One key innovation Stripe introduces is shared payment tokens, which allow agents to securely share payment credentials with sellers while enforcing strict limits on amount, currency, time, and seller scope. This mechanism prevents overspending or misuse of credentials, even if the agent misinterprets pricing or is duped by fraudulent sites. Kaliski demonstrates this with a live example where a shared payment token restricts a Visa card to a $25 limit for a specific seller, and Stripe enforces this limit during payment attempts, ensuring controlled and auditable transactions.
Beyond credential sharing, Stripe collaborates with partners like Tempo to develop protocols such as the machine payments protocol, enabling agents to pay for API calls or tool usage via HTTP requests that can signal payment requirements and accept payments programmatically. Additionally, the Agent to Commerce Protocol (ACP), developed with OpenAI, standardizes how agents interact with e-commerce checkouts by exchanging structured data about products, pricing, taxes, and shipping. This protocol ensures that agents can reliably and deterministically complete purchases without relying on fragile web scraping or UI interactions, reducing disputes and chargebacks.
In conclusion, Kaliski stresses the importance of maintaining non-determinism for discovery and planning while enforcing determinism for payments and checkout processes. Stripe’s solutions, including shared payment tokens, machine payments protocol, and ACP, aim to create a safe, flexible, and agent-friendly payment ecosystem that supports various payment methods and preserves seller control and risk management. This infrastructure is designed to enable the autonomous economy where robots can transact securely and efficiently with businesses, minimizing financial risks and maximizing trust.
