What is agentic commerce, actually?

The one-sentence version

Agentic commerce is what happens when the customer on your store isn't a person — it's an AI agent acting for one.

Not browsing. Transacting: searching the catalog, comparing variants, checking stock, assembling a cart, sometimes checking out. The shift that matters isn't that AI can describe products — it's that software can now do the shopping.

Agents that browse vs agents that transact

Most 'AI commerce' today is browsing-grade: an assistant reads your product page and summarizes it for a human. Useful, but the page was built for eyes — the agent is scraping layout and hoping.

Transaction-grade is different. The agent calls a tool: search_catalog('walnut dining table under $2000') and gets structured results — names, variant IDs, live prices, real stock. It doesn't parse your theme. It asks your store questions and gets answers it can act on.

That difference is the entire ballgame. A browsing agent can mention your product. A transacting agent can sell it.

What a store needs to expose

For an agent to transact, your store needs to answer machine questions as a service: catalog search, product detail, variant availability, pricing, shipping options, cart actions. In practice that's a typed API — JSON-RPC tool endpoints, increasingly spoken through MCP (Model Context Protocol), which is becoming the standard connector between agents and the tools they call.

Underneath the API sits the unglamorous part: the catalog has to be correct. An agent that reads stale stock or a drifting price doesn't forgive the way a human squinting at a page does — it acts on bad data at machine speed.

This is why agentic commerce is an operations story wearing an AI costume. The stores that win it are the ones whose product data is already trustworthy enough to expose directly.

Is your store agent-readable?

A five-question checklist, in rough order of what breaks first:

One: can a machine find your products without parsing HTML? An endpoint, a feed, a structured catalog — anything but 'read the theme.'

Two: are price and stock accurate at read time, not nightly-batch time?

Three: do variants resolve to purchasable entities — real SKUs, real options — not display text?

Four: is there a tool surface beyond read-only? Search and detail are table stakes; cart and checkout actions are where transactions happen.

Five: can you see it working? The honest test isn't a deck — it's calling the endpoint yourself. Our own MCP server answers tools/list publicly at tendrils.io/mcp, today, from any agent. If a vendor claims agentic readiness, ask for the endpoint.

Where this is going

The optimistic read is that agents become a real sales channel — tireless, data-literal buyers who comparison-shop at machine speed. The sobering read is the same fact: a channel that never browses your branding, never sees the lifestyle photography, and buys purely on catalog truth.

Either way, the work is identical: correct data, exposed as tools, on a store that can prove it runs.