Retailers running shopping agents on Claude have seen carts up to 35 percent larger and shoppers 60 percent more likely to complete a purchase, according to Anthropic’s own figures from its retail partners. On September 3, 2026, Anthropic decided every business should be able to test that claim for itself. The company open sourced Claude Commerce Agents 2026, a full blueprint for building shopping and merchant AI agents, complete with runnable code across four industries.

This is not a new model or a hosted product. It is documentation, harness code, and prompting patterns meant to compress months of agent architecture trial and error into a repository anyone can fork. For entrepreneurs and developers who have wanted to build agentic commerce experiences but did not know where to start, this release matters. Here is what is actually inside it, how it works, and what to watch before you build on top of it.

What’s Inside the Agentic Commerce Architecture

The repository, published at anthropics/commerce-agents under an Apache 2.0 license, contains two distinct agent harnesses rather than one generic chatbot template. The shopping agent lives inside a merchant’s own app. It searches the catalog, handles multi-item requests, compares options against constraints like price or delivery date, builds the cart, and answers order and returns questions in the same conversation.

The merchant agent serves store staff instead of shoppers. It answers sales performance questions, flags inventory issues, recommends pricing and promotions, and drafts marketing campaigns. Both agents ship with four runnable vertical demos: retail, travel, telecom, and entertainment, showing the same underlying harness adapted to different catalog shapes and transaction types.

Everything runs locally on Python 3.11 or newer and Node 22 with an Anthropic API key, and the code deploys the same way across the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI. A companion Claude Code plugin called commerce-builder scaffolds a new agent against a team’s real backend rather than a synthetic demo store, which is what separates this from a typical proof of concept repo.

Merchant AI Agents and the Skills Versus Subagents Debate

The most transferable idea in this release has nothing to do with commerce specifically. Anthropic argues against routing a shopping conversation through separate subagents for each domain, such as one for search and another for checkout. A commerce session is one tightly coupled conversation, and every handoff between subagents is state lossy: the orchestrator has to reconstruct the cart, preferences, and history each time, which costs extra tokens and adds seconds of latency.

Instead, the blueprint uses agent skills, modular instructions that load into the same agent that already holds the conversation history. Anthropic reports that a single agent using skills beat both a one giant prompt design and a subagent design on quality, often at lower cost and latency, across several enterprise deployments. You can see a similar pattern play out in our breakdown of the best AI agent frameworks for production deployments, where orchestration choices matter as much as model choice. Subagents still earn a place for narrow, self-contained work like deep research, just not for the core commerce loop.

The shopping and merchant agents also carry different guardrail risks. A shopping agent’s failure mode is recommending the wrong item or over-persuading a buyer. A merchant agent’s failure mode is giving an out-of-policy answer or hallucinating stock levels. Treating these as separate harnesses with separate evaluation criteria, rather than one undifferentiated commerce bot, is the more interesting engineering decision here.

How Businesses Can Actually Use This Blueprint

If you are wondering how do AI shopping agents work well enough to trust in production, the honest answer from this release is: caching, streaming, and code level guardrails matter as much as the model. Anthropic’s engineering notes point to prompt caching as the main cost lever, with cache reads costing roughly a tenth of fresh tokens and the best deployments hitting 90 to 99 percent cache hit rates by ordering prompts from global rules down to session specific and volatile data. This complements the persistent memory techniques covered in our guide to how AI agent memory works, since the blueprint runs memory extraction asynchronously for better recall.

For latency, the blueprint uses eager tool dispatch, executing each tool call as its arguments finish streaming rather than waiting for the full response, which reportedly cuts multi-second gaps down to a few hundred milliseconds. UI elements like product cards and itinerary comparisons are implemented as typed tools rather than custom text the model has to format, so conversation history stays native and reloadable.

Critically, money movement, database writes, and order IDs are gated in code, not left to the model’s judgment. The agent proposes an action and the harness applies it. For a small business or product team evaluating whether to build a shopping assistant in house, this is the practical starting checklist: skills over subagents for the core loop, aggressive prompt caching, typed tools for anything visual, and hard guardrails around anything that touches money.

What to Watch Before You Build on It

The 35 percent and 60 percent figures deserve a caveat that Anthropic itself acknowledges: they come from its own retail partners, are not independently audited, and “up to” language means they represent best case results, not averages. A shopping agent bolted onto a messy, poorly instrumented product catalog will not automatically reproduce them, a lesson that echoes what we found covering Google’s own AI shopping agents rollout earlier this year.

This release also raises a genuine architectural question for the agentic commerce field. Anthropic bet on vertical specific reference implementations rather than one universal commerce agent, arguing that a travel itinerary with cancellation policies has almost nothing in common with a telecom billing and contract flow. Whether teams actually reach for the closest vertical demo, or end up needing a fifth or sixth vertical Anthropic has not built, is the open question worth watching as adoption spreads through 2026.

The Takeaway for AI Agent Builders

Claude Commerce Agents packages real lessons from retail deployments into a forkable starting point instead of a black box. Three things matter most: the skills over subagents architecture is a lesson that applies well beyond commerce, the reported conversion gains are directional rather than guaranteed, and the harness still requires real integration work against your own inventory and payment systems before it earns any of those numbers.

If you are building or evaluating AI agents for your business, explore more breakdowns of agentic commerce, agent architecture, and enterprise AI tools at BigAIAgent.tech. For the full technical details, see Anthropic’s own Claude Commerce Agents announcement and the open source repository on GitHub. Would a documented, forkable blueprint like this change how your team approaches building an AI shopping or support agent, or does your business still need a fully managed platform instead?