Observed facts: LangChain describes Restock as a Slack agent that searches products and places retailer orders through Zinc’s MPP API. The user reviews the cart in Slack and separately approves the payment in Stripe Link. In the illustrated order, $23 was approved, the charge was $22.18, and $0.82 was refunded. LangChain says it ran a hosted deployment through to order_placed and received the expected refund.
Mechanism: MPP supplies the payment exchange over HTTP: the merchant returns a 402 challenge with payment instructions, and the client retries with a payment credential. Link supplies wallet consent. The payment tool retrieves a Link token from a private sandbox file, deletes it after use, and checks the order against the reviewed cart and approval freshness. As LangChain describes it, the model does not receive payment secrets.
Why now: As agents move from product discovery toward transactions, payment credentials, irreversible spending and approval boundaries become operational design questions. This example offers a specific, inspectable architecture rather than only a claim that agents can buy things.
Implications: The implementation suggests that agent commerce can be structured around discrete authorization steps and bounded amounts. Its practical reach, however, depends on the availability and reliability of MPP-enabled merchant APIs.
Evidence boundary: These implementation details and the live-order claim come from LangChain’s own blog post; the stated prices are illustrative, and the supplied evidence does not independently verify the deployment or establish broad retailer coverage.
Watch condition: Watch whether MPP-enabled retail APIs expand and whether independent deployments reproduce the approval, credential-isolation and refund controls.