Anthropic published an open reference implementation for building commerce agents on Claude: a shopping-agent starting point, sector examples (retail, travel and others), and an engineering guide. Klaviyo’s 8 September write-up is useful because it separates the repo from production.

Anthropic’s own framing (as Klaviyo quotes it): an open reference implementation, not a supported product. No SLA. It shows how an agent should call search, cart, inventory and pricing systems. It does not arrive knowing your customers, policies or which actions need a human.

Klaviyo’s build-vs-buy checklist is four requirements. Hold any DIY fork or vendor to them:

  1. Real customer context (which order, VIP vs first-time, open tickets, what the brand may do next).
  2. Working integrations with authority rules (orders, returns, loyalty, help desk: access is not the same as permission to act).
  3. Channel-specific UX (web chat cards vs short SMS/WhatsApp replies; language switching).
  4. An improvement loop someone owns after launch (bad answers, handoffs, model cost, regression tests).

Klaviyo estimates a broad production build as a multi-quarter engineering programme (overlapping design, integrations, security, guardrails and testing). Treat those week ranges as vendor framing, not your Gantt. The same post pitches Klaviyo Customer Agent and cites Naked Wardrobe (86% autonomous resolve / 94% on product recommendations over 90 days). That is a self-reported case study. Useful as a reference point; not audited proof for your catalogue.

For EU ecommerce solos and Spanish agencies, the expensive mistake is “fork Claude, ship agent” before GDPR/AI Act chatbot disclosure, WhatsApp/SMS rules, and handoff policies are designed. The free repo does not include on-call, DPIA, or who is liable when the agent refunds the wrong order.

Build yourself when the agent is the product advantage (proprietary search, pricing logic, non-standard catalogue). Otherwise compare total cost of owning integrations and the improvement loop against a ready agent that still lets you control knowledge, skills and permissions.

What to do this week: write the four requirements against your stack; estimate who owns on-call; decide build vs buy before anyone clones the blueprint into production.