The reference implementation makes it easier to start. The harder work is turning it into a dependable customer experience.
This week, Anthropic published a blueprint for building commerce agents on Claude. The release includes a reference shopping agent, examples for retail, travel, telecom, and entertainment, and an engineering guide that shows how the pieces fit together. Klaviyo is featured alongside Shopify, Visa, Mastercard, and Square, building on its existing integration with Claude.
At first glance, the path from repository to production looks short: fork the code, connect your catalog and storefront, and start testing. That's a reasonable read, but it only covers the easy half of the work.
The short version: a demo can run in minutes. Getting that agent into production can take an engineering team two or more quarters.
The real difference lies in the work that begins once real customers, live business systems, and brand-specific decisions enter the picture. So what actually stands between a demo and a dependable experience?
What a blueprint is and what it isn't
Anthropic is clear about what it released: “an open reference implementation, not a supported product.” It comes without a service-level agreement, and your team must connect it to the systems you run. Those boundaries are intentional. The blueprint deliberately stops where your business starts.
Its guide shows how an agent should call your existing search, cart, inventory, and pricing systems, and how to structure, control, and test its behavior. It does not connect or operate those systems or arrive already knowing your customers, policies, and operations.
That gap is where most of the production work, and most of the customer experience, lives.
What an agent actually needs
4 requirements determine whether an agent can move beyond answering questions to serving customers reliably. Whether you build or buy, hold your team or vendor to them.
1. Context the agent starts with
A customer asks, “Where is my order?” It sounds like one question, but answering it well requires the agent to make sense of several things at once: which order the customer means, whether it has shipped, whether it is late, whether another service issue is already open, and what the brand is allowed to do next. The right response may also depend on whether this is a first-time buyer or a long-time customer.
An agent that can see that context answers differently from one that can only retrieve a tracking number. It needs the customer data your brand already has, along with your policies, product knowledge, and rules for which questions it can handle and which ones need a person. Without that foundation, the response may sound fluent while still feeling generic.
2. Integrations that already work
A production agent needs access to the systems behind the journeys it handles, including order management, returns, subscriptions, loyalty, and the help desk. A blueprint shows you where those connections go. Your team still has to build them, keep them working, and decide what the agent is allowed to do through them.
Consider a customer who says, “Redeem my points.” The agent has to do more than find a balance. It needs to know who is asking, have permission to act, complete the request, and respond appropriately if something goes wrong. Access is not the same as authority. A production agent needs clear rules for what it can do, what requires confirmation, and when a person should step in. The higher the stakes, the clearer those rules need to be.
3. An experience designed for each channel
Customers experience one brand. Each channel still has its own rules. An answer that works in web chat may be too long for SMS. A product recommendation that works as a carousel with images and a button to shop needs a different treatment when it becomes a text message.
The hard part is designing how the agent responds, asks follow-up questions, presents recommendations, completes actions, and hands off in each channel. For global brands, that work also extends across languages. When you build it yourself, adding another channel means designing, testing, and maintaining another customer experience.
4. Someone to run the improvement loop
The improvement loop is the part teams underestimate most because it rarely appears on a feature list. An agent isn’t finished on launch day. Someone has to determine whether a problem came from missing knowledge, a bad handoff, an overly strict rule, or the model itself, then fix it without creating another issue.
Your team also has to compare models, decide when to update them, and confirm that each change improves accuracy and quality without driving up cost. That cycle continues for as long as the agent is live, making it part of both the cost and the operating model.
Together, these 4 requirements change the question from whether your team can build a demo to whether your brand wants to own the production system behind it.
A faster path to production
Buying changes what your team has to own. With an out-of-the-box agent, the platform is already built and supported, while your team retains control over the brand knowledge, policies, permissions, and outcomes that shape the experience.
Klaviyo Customer Agent is one example. Here is how it addresses the same 4 requirements:
1. The context is already there, and it keeps getting richer.
For brands already using Klaviyo, Customer Agent can draw on the order history and preferences already available in Klaviyo, together with the brand content, guidance, and policies your team configures. It also adds useful signals back to Klaviyo: product recommendations are recorded as events, and custom skills can save confirmed preferences to the customer profile.
Those signals inform future conversations, segments, and messages. Klaviyo’s third-party LLM providers are contractually prohibited from using customer data to train their models.
2. Supported integrations are ready to connect.
Connect a supported ecommerce platform or loyalty, subscription, returns, or customer support partner in Klaviyo, and Customer Agent can use the data and actions that integration makes available. For those supported journeys, your team does not have to build and maintain the integration itself.
3. Configure one agent, then turn on each channel.
Configure Customer Agent with your catalog, brand knowledge, skills, and integrations, then enable web chat, email, SMS, or WhatsApp as each channel is ready. Klaviyo adapts the experience to the channel, from rich product and order cards in web chat to concise replies and deep links in messaging. It also detects the language of each message and responds automatically across 113 languages, even if the shopper switches languages.
4. Klaviyo runs the improvement loop with you.
Klaviyo maintains the agent platform, supported integrations, and pre-built skills. It runs benchmark tests, selects and updates the models behind the agent, and optimizes response quality, accuracy, and cost over time. Your team focuses on refining brand knowledge, policies, permissions, and handoff rules, measuring business outcomes, and adding custom skills as your needs evolve. You decide what good looks like without having to build and operate the system that keeps getting you there.
Naked Wardrobe adopted Customer Agent at launch. Over the next 90 days, Customer Agent autonomously resolved 86% of its customer queries and 94% of its product-recommendation queries.
What it takes to build your own agent
A focused pilot that searches a catalog and answers a narrow set of product questions can move quickly. A production agent that recognizes customers, takes actions, works across channels, and hands off safely takes longer.
Based on the scope of Anthropic’s blueprint, we estimate the following ranges for a broad production build:
- Design and prototype: two to 4 weeks. Choose the first customer journeys, connect them to your real catalog, and define what a good answer looks like.
- Integrations: 6 to 12 weeks. Connect the catalog, cart, customer profile, order, policy, and fulfillment systems needed for the experience you want to launch.
- Identity and security: 3 to 6 weeks. Add customer authentication, permissions, credential handling, and abuse controls around the reference implementation.
- Guardrails and approvals: 4 to 8 weeks. Decide what the agent can do, what requires confirmation, and when a person should take over.
- Testing: 6 to 10 weeks. The repository includes software tests and a workflow for creating deployment-specific test cases, but your brand still has to supply the catalog, policies, edge cases, and real customer scenarios. Anthropic recommends starting with 50 to 100 cases per user flow. Your team also needs to test each flow in every channel where it will run.
These workstreams overlap, so the ranges are not additive. Even so, a broad launch can become a multi-quarter program involving engineering and product, with input from legal, customer experience, and merchandising. That gets you to launch, not the finish line, because the improvement loop begins as soon as real customers arrive.
Build or buy? Look beyond the demo
Anthropic has made it easier to reach a working demo. That speed can make the work after the demo easy to underestimate. A team that forks the blueprint still owns integrations, testing, security, on-call support, channel changes, and continuous improvement. Compare the total cost of that commitment with the cost of an out-of-the-box product. The starting code may be free, but that doesn’t tell you which path will cost less.
For a small number of brands, owning that work is the right choice. Here is the test I would use: build when the agent itself is part of your product advantage. That may be true if you own proprietary search, ranking, pricing, or transaction logic, sell something that does not fit a standard product grid, or have genuinely specialized customer conversations. Prove that need through testing before committing.
For most brands, including many enterprise teams with the capacity to build, an out-of-the-box agent is the better place to start. Choose one that is ready to run without boxing your brand in: custom skills, grounding in proprietary knowledge, a credible path to the integrations you need, and controls that let your team refine responses and improve performance over time. That keeps engineering focused on what differentiates your brand.
If you’re looking for an agent that’s ready to run, take a look at Klaviyo Customer Agent.
Related content
