Teaching agents to pay — Anna Spysz, Stripe
Summary
This talk introduces 'agent-to-commerce,' detailing how AI agents can move beyond research assistance to execute complex transactions on behalf of users. The infrastructure relies heavily on structured data protocols like the Universal Commerce Protocol (UCP) and requires merchants to publish a `merchant capabilities manifest` and JSON catalogs, rather than relying on traditional website browsing. Security is maintained through shared payment tokens, ensuring that the payment provider—not the agent or merchant—enforces transaction limits and guardrails.
Key takeaways
-
Agent Discovery vs. Human Browsing
5:13
Agents do not browse websites; they read structured data (JSON) from a `merchant capabilities manifest` to understand supported payment methods, endpoints, and product catalogs. This makes merchants 'agent-ready.'
-
The Role of UCP in Transactions
5:40
The Universal Commerce Protocol (UCP) acts as a shared language defining how agents initiate, update, complete, and cancel purchases across multiple merchants and agents.
-
Security via Shared Payment Tokens
13:46
During checkout, the agent receives and passes a 'shared payment token' (representing a raw card number or wallet). The seller only unwraps what is necessary, and the payment provider enforces all limits, mitigating risk.
Technical details
-
Agent Architecture
730s
An agent's decision-making process is structured: a Large Language Model (LLM) 'brain' makes decisions, which are executed by defined tools/actions. These actions are governed by instructions that run in a loop and shaped by the system prompt (persona and ethics policy).
-
Merchant Integration Requirements
630s
For agent access, merchants must provide a publicly accessible JSON file—the `merchant capabilities manifest`—located in the `.well-known` directory. This manifest declares supported payment methods and API endpoints.
-
Data Structuring for Agents
645s
Product catalogs, policies (e.g., shipping/return), and descriptions must be provided as structured JSON data to allow agents to filter and justify recommendations accurately.
-
Agent Guardrails Checklist
750s
Critical guardrail checklist items for building agentic commerce include: always disclosing that the user is speaking to an AI, upfront disclosure of fees, respecting 'stop' or 'cancel' commands, capping totals at a user-defined maximum amount, and logging all decisions for auditability.
Mentioned resources
- Stripe Developers
- Stripe Developers YouTube Channel
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.