What is agentic commerce?

“Agentic commerce” is the phrase people reach for when the thing buying or selling isn’t a person clicking a mouse — it’s a software agent acting on someone’s behalf. Strip away the hype and the idea is narrow and concrete: a store that a machine can read, reason about, and transact with as reliably as a human can.

That’s it. No sentience, no science fiction. An LLM with a shopping task — “reorder the filament I bought last month, but only if it’s still under $30” — needs to find the product, check the price, confirm it’s the right one, and place the order. Today it does that by scraping a page built for human eyes, guessing which button is “Add to cart,” and hoping the checkout flow doesn’t change next week. Agentic commerce is the bet that stores will stop making agents guess.

Two kinds of agent, two problems

It helps to split the term in half, because the two sides have almost nothing in common technically.

Buyer agents act for a shopper. They browse, compare, add to cart, and check out. The hard part here is trust boundaries: the agent is authenticated as a real customer, so it can see that customer’s addresses and order history — and it must not be able to do anything the customer couldn’t do themselves. A buyer agent that can read your saved cards but not spend beyond a limit you set is a different security posture than a human session.

Merchant agents act for the store. They write product descriptions, adjust pricing, triage returns, reconcile inventory against an ERP. The hard part here is scope: a merchant agent should be able to publish a banner but maybe not delete a category; edit a price but within guardrails; issue a refund but leave an audit trail. That’s role-based access control, just pointed at a bot instead of a staff login.

Why bolting an agent onto a normal store is hard

Most e-commerce platforms were built for one consumer: a browser rendering HTML. Everything downstream assumes a human is in the loop to notice when something looks wrong. Agents remove the human, and all the soft assumptions get sharp:

• Idempotency. A human rarely double-submits an order. An agent retrying a timed-out request will — unless every write is keyed so the second attempt is a no-op instead of a second charge.
• Structured contracts. Screen-scraping breaks the moment a designer moves a div. Agents need typed, versioned responses that don’t change shape when the CSS does.
• Auditability. When a bot makes a mistake at 3 a.m., you need a log that says exactly what it did and lets you reverse it. Human mistakes are self-limiting; automated ones scale.

None of these are exotic. They’re the same disciplines good API teams have practiced for a decade. What’s new is needing them at the storefront, not just in a back-office integration.

The honest part: you may not need this yet

If no agent is transacting against your store today, agentic commerce is a roadmap item, not an emergency. Plenty of businesses will do fine for years with a conventional platform and a human checkout. We’d rather say that plainly than pretend the sky is falling.

But the direction of travel is clear enough that building the plumbing now — structured APIs, idempotency keys, an audit log — costs little and ages well, because those same properties make your store easier to integrate, test, and debug for humans too. The agent-readiness work is mostly just good engineering with a sharper deadline.

How Wardenclyffe approaches it

The Wardenclyffe AI Engine treats the agent surface as a first-class part of the store rather than a plugin. Every action the human UI can perform is also exposed as a tool over the Model Context Protocol, with the same authentication, the same role checks, the same audit trail, and the same idempotency keys. There are two endpoints — one for buyer agents, one for merchant agents — so the two trust models above stay cleanly separated.

Because the whole thing is source-available and self-hosted, you can read exactly how the agent boundary is enforced instead of trusting a vendor’s marketing. And if you’d rather build your own agent layer on top of a typed backend, that’s a supported path too — see our take on vibe coding a real store on .NET for where that works and where it bites.

Agentic commerce, in the end, isn’t a product category. It’s a property a store either has or doesn’t: can a machine use you without guessing? Everything else is detail.