Agent commerce needs a confirmed purchase and a costed outcome
This week's Runtime and cost updates show where a commerce agent's tools end and a business decision begins.
DataInbox
Architecture & Research
Shopify's October 13 Shop MCP release changes what a commerce agent can do: one MCP connection can search its catalog, build a cart and complete checkout after the shopper confirms the total and payment method. A signed-in shopper can also use saved Shop Pay details and find orders. This is a released Shopify capability, not a DataInbox integration.
We expanded the existing [Agent Runtime explanation](/agent-runtime) with a concrete commerce case. The service owner needs to know whether the shopper approved the actual total, whether checkout completed, and which merchant order now owns the result. Shop MCP places orders in the merchant's Shopify admin. At launch there are no webhooks, so an agent must poll checkout status. A proposed DataInbox Companion could keep the service request, permitted context, approval boundary and returned status together, then reconcile an uncertain outcome before a follow-up action. That is an architecture hypothesis to test, not shipped DataInbox functionality.
Before: a shopper asks an agent about availability, leaves for checkout, and support later pieces together whether an order exists. After, in a bounded pilot, the agent could prepare the cart while the shopper confirms payment through Shopify; an Inbox could retain the request and reconcile the returned order status. The merchant remains the order authority. Test completed purchases, abandoned or corrected carts, support handoffs, time to a complete answer, approval and polling delay, and the total cost per successful outcome. The pilot needs current stock and price checks, identity and consent, provider payment handling, and recovery for uncertain status. Shopify's announcement does not prove a conversion gain or lower support cost.
Anthropic's October 14 beta OpenTelemetry export is a related evidence development: teams can trace Messages API requests and Managed Agents sessions in their own observability backend. It can help investigate a provider call, while the business record still needs to say who authorised the operation and what result was accepted. The export is beta; it does not create DataInbox evidence automatically.
Anthropic's October 14 release note
Price the complete path
We also refreshed the existing [AI Economics Dashboard](/labs/ai-economics) and [AI Cost Benchmark Lab](/labs/ai-cost-benchmarks). The eleven catalogued models' standard USD input, cached-input and output tariffs remain unchanged in this check. The latest available ECB reference rate at our October 15 check was October 14: EUR 1 = USD 1.1195. Only current euro conversions changed. The default 380-use Terra example is still $59.22 of estimated provider tokens, now about €52.90 instead of €52.96. Those numbers are illustrative, not a measured DataInbox saving.
The dashboard's stored DataInbox Runtime team scenarios were calculated on August 24 and were not rerun. The token comparison holds 1,000 questions at 10,000 input and 500 output tokens each; equal token counts do not establish equal quality. Payment, tools, cache writes and storage, grounding, faster or regional processing, tax, review and errors can add cost. A useful pilot should compare a completed, correctly authorised outcome and its full operating cost, not just the model token price.
The locally verified DataInbox CLI is rc.50 with a matching command contract and skill. It supports reviewing and validating intended configuration changes offline, but local commands alone do not prove availability on a particular server or Inbox. We made no new product release claim from that version check.