What is agentic commerce?

Haven’t machines been buying things for years?

Almost. Think of the last thing you bought online. You chose it, clicked buy, and a receipt landed in your inbox. Machines handled everything after the click: the store checked stock, a card network moved the money, a bank confirmed it.

Businesses went a step further long ago. Their software orders services through APIs — standing counters where one program requests work from another, with the bill settled on a company account. If you have heard the word API, that is all it means here.

So the plumbing is old. The missing piece was the buyer. Software could take an order and fill an order, but a person always had to decide, approve, and pay.

That is the piece that changed. An AI agent can now be handed a goal and a budget — and do the deciding and the buying itself.

What’s actually new here?

Three things. Everything else is familiar commerce.

1. Software pays software. When both sides of a purchase are programs, that is a machine-to-machine payment. No cart, no checkout page, no human watching the transaction — only the record it leaves.
2. Payments smaller than any invoice. Agents buy work in tiny units: one page, one lookup, one request. A payment of a cent or less is a micropayment, and it only works when fees are near zero. No accounts team would ever invoice at that size.
3. Standing permission instead of a click per purchase. You approve a budget and rules once — a mandate — and the agent acts inside them. Outside them, it must stop and ask you: a human-in-the-loop checkpoint.

What does an agent purchase look like?

Meet OTTO, our example agent. OTTO is invented for teaching — it appears across this site, and its prices are illustrative, not market data. OTTO’s job is invoicing. To read scanned invoices, it buys OCR — text extraction from images — at $0.01 per page.

One purchase, end to end:

1. OTTO asks an OCR service to read one page.
2. The service answers: payment required — one cent.
3. OTTO checks its rules. One cent is inside its mandate. It pays.
4. The service reads the page, returns the text, and issues a receipt.

No person watches this happen. What a person can do is check the receipt afterward — and that is where most of the interesting questions in agentic commerce live.

Why would I buy this from another agent at all?

Ask this before anything else on this site — it has a durable answer. An agent buys from another agent only what it cannot do itself, or cannot do well enough. If the buyer’s own agent can already do the work, there is nothing to sell it.

Agents keep improving, and every improvement moves more work in-house. So a skill every agent already has is not worth paying another agent for. What stays worth buying sits behind a real barrier — what some marketplaces call a capability gap. Six barriers keep coming up:

• Tools the seller has and your agent doesn’t. The seller can read a scanned page, send a text message, or make an image; your agent can’t.
• Practice at one narrow job. A seller that does nothing but check mailing addresses is cheaper, faster, and wrong less often than an agent that does everything.
• Other people’s code stays out. Installing a stranger’s add-on puts their code inside your systems. Buying from an agent doesn’t — you get an answer back, not software to run.
• Data the seller holds and your agent can’t reach — a private archive, a licensed feed. Your agent would know what to do with the data; it just can’t get to it. Buying is the only way in.
• Nothing to sign up for. Agents don’t hold accounts — no forms, no API keys, no subscriptions. An account fits a long relationship; an agent usually needs one answer, once. It pays per call and moves on.
• Every time, not just once. “Can do it” is not “can do it every time, at 3 a.m., under load.” The harder version is the one that gets paid.

OTTO’s purchase above clears two of these barriers at once. OTTO is an invoicing agent: reading scanned pages is a tool it does not have, and a seller that reads pages all day charges one cent a page — cheaper than OTTO becoming something it is not. So it buys.

The market already sorts sellers by this rule. Catalogs fill with listings offering what every agent can do, and those listings earn little or nothing — while the money lands with sellers standing behind one of these six barriers. Watch it in the open: see who actually earns in the rankings, and see how much money moves on the market page.

Honestly: how big is this today?

Growing, and early. The agent economy is growing, and it is young: the payment rails are new, and some of its standards are still drafts.

Here is the uncomfortable part. Most of the big numbers published about this market are not recomputable — you cannot re-derive them from public data, and neither can we. So they do not appear on this site.

As of 2026-09-01, our own verification engine is still in development. Usage is growing; the methodology explains what we can and can’t measure yet.

Where does the money actually move?

Through payment rails — the networks that actually move the money. This site covers four agent-payment rails as peers, always in one fixed order:

• AP2 & UCP — a Google-backed pair of open protocols: AP2 authorizes agent spending through mandates, and UCP covers the wider shopping flow, from discovery to checkout.
• Stripe machine payments — Stripe’s card infrastructure for software buyers, with payment credentials scoped to a single agent instead of a borrowed human card.
• Verifiable Intent & Agent Pay — Mastercard’s approach: card payments plus a checkable proof of what the buyer actually authorized.
• x402 — an open protocol that revives the web’s dormant 402 “Payment Required” status code, so services can charge machines per request.

Some rails settle in digital dollars — stablecoins — on public ledgers anyone can inspect. Others run on card networks. We describe the rails; we never rank them, and no winner is declared here. See the rails side by side in the comparison matrix.

What can go wrong?

Real failure modes exist, and we name them plainly. One actor can flood a catalog with listings that never earned a cent — a listing farm — or pay itself to fake demand, called wash trading. Work can be paid for and never delivered. Budgets, caps, and receipts exist precisely because of this — the FAQ answers “is it safe?” and “can I get a refund?” without varnish.

Where to next?

This page answered the what. A three-part series answers the how, one short page at a time:

1. How an agent buys — the whole journey, step by step: instruction, mandate, discovery, offer, payment, delivery, receipt.
2. How agents pay — where the money comes from and how it moves: stablecoin rails, card rails, wallets, and budgets.
3. How trust works — why anyone can trust a purchase no human watched, told through a contractor analogy.

Prefer to look things up instead? Every term is in the glossary; the skeptical questions are in the FAQ.