Descriptions:
Steve Kaliski, principal software engineer at Stripe, takes the AI Engineer stage to address one of the least-discussed infrastructure problems in the agentic AI stack: how do you let autonomous agents spend money safely, and how do businesses accept payment from non-human buyers?
Kaliski frames the core challenge around four failure modes — wrong merchant, wrong item, wrong amount, wrong credential — and explains why the naive approach of handing an agent a credit card number and a browser is fundamentally broken: it’s slow, hard to observe, and has an unbounded blast radius. His central thesis is that discovery and exploration can tolerate non-determinism, but credentials, payments, and checkout require it — and that separation drives Stripe’s entire design approach for agents.
The talk then walks through three concrete solutions Stripe and its partners have shipped. First, a credential-sharing system that issues scoped payment tokens with hard spend limits (demonstrated at $25) and expiry windows (30 days), with live demos showing a $50 charge being rejected by Stripe’s enforcement layer. Second, the Machine Payments Protocol — developed with Tempo — which uses HTTP 402 status codes to signal payment requirements within API tool calls, enabling agents to pay for tool access programmatically. Third, an API-driven checkout flow that replaces browser form-filling with verifiable, deterministic transactions. Engineers building financially-capable agents on Stripe’s infrastructure will find this talk a practical reference for the full payment stack.
📺 Source: AI Engineer · Published June 06, 2026
🏷️ Format: Keynote Launch
