Anthropic just open-sourced the commerce-agent pattern we've been hand-building

Anthropic published reference code for two agent types this month: a shopping agent for customer-facing search and cart help, and a merchant agent for staff-facing analytics, listings and campaigns. Same prompt, same skills-and-tools contract, three ways to run it (Messages API, Agent SDK, or Claude-hosted). The part worth your attention isn't the demo. It's the safety layer baked into the code, not the prompt. Every write the merchant agent wants to make, a price change, a restock, a campaign draft, gets staged and held until a person clicks approve. Server-issued IDs only. One write in flight at a time. That's the exact draft-then-review pattern most of us have been wiring up ourselves, one integration at a time, except now there's a public reference implementation for it. If you're running a commerce stack (Shopify, a headless storefront, HubSpot for the B2B side), this is worth an hour of your time even if you never deploy it as-is. The backend interface pattern, you implement StorefrontBackend or MerchantBackend and the model calls into it server-side, is the cleanest illustration I've seen of "the agent proposes, your code disposes." Repo: https://github.com/anthropics/commerce-agents What's the equivalent approval gate you've built by hand for your stack? I'm curious how many of us independently landed on the same staged-write pattern before this existed.