Not a thought piece. It's live: as of this month, a piece of software with a funded wallet can buy our teaching materials from this website in about two seconds, with no human involved on either side. Here's the full build — what it cost, what broke, and the honest answer to "has anyone actually bought anything?"
A growing share of the research that used to happen in a Google search box now happens inside an AI assistant. A teacher asks their assistant to find AI professional development that isn't rubbish. A manager asks theirs to shortlist automation training. The assistant goes and looks — and what it finds is websites built entirely for human eyeballs: hero images, testimonials, buttons. Nothing it can reliably parse, compare, or act on.
Our bet is simple: the next version of SEO is being legible to agents. And the step after legible is transactable. If an agent shows up doing genuine research on behalf of a human with a budget, we want two things to be true: it can read our catalogue as structured data, and — if a $1 artefact would help its human right now — it can just buy it, instead of filing our contact form into a report nobody reads.
The protocol that makes the second part possible is called x402. It resurrects HTTP status code 402 — "Payment Required," reserved in the web spec since the 1990s and never used. The server quotes an exact price in the response; the agent pays in USDC (a digital dollar) on the Base network; the content comes back in the next response. No accounts, no cards, no checkout page, no payment processor taking 30 cents plus 2.9%.
The refusal is the interesting part. A normal website says "payment required" with a checkout page a machine can't operate. An x402 response says it in a format any compliant client can act on without a human:
The agent's client library reads that, signs a USDC transfer authorisation with its wallet, and repeats the request with the signed payment in a header. We verify the signature, settle the transfer on-chain, and return the content — all in the same request-response cycle. The whole sale takes about two seconds. The money lands in the company wallet. There is no invoice, no reconciliation, no "your payment is being processed."
The pricing rule we settled on: price like an API, not like a course. An agent doing research for its human isn't going to authorise a full course enrolment — but 25 cents to $5 for a structured artefact that directly answers the brief? That's a rounding error on the errand it's running.
And one endpoint that costs nothing, which is secretly the most important one: a free machine-readable catalogue of everything we do — including our human courses, with prices and enrolment links. That's the agent-SEO asset. Most agents that find us won't buy an artefact; they'll relay "here's a Perth outfit with teacher PD, here's the price, here's the link" to their human. The free endpoint makes that relay accurate instead of hallucinated.
Our website is deliberately boring: static HTML on Vercel, no framework, no build step. The store didn't change that. Five steps:
One file, about 200 lines of plain Node, deployed as serverless functions beside the static site. It does the whole dance: return the 402 with payment requirements, and when a payment header arrives, verify and settle it before serving content. No npm packages in production. If you can write an HTTP handler, you can write this.
Before any paid endpoint existed, we shipped the free one. Structured JSON: who we are, every course, every price, every enrolment link. If you build only one thing from this article, build this — it's the part that pays off even if agent payments take years to matter.
The paid material lives inside the serverless functions themselves, not as files on the website. There's no URL that serves it unpaid, nothing for a crawler to scrape around the paywall. The sources stay in the repo; a small build script regenerates the bundled versions when they change.
Verifying signatures and settling on-chain is the one genuinely fiddly bit, and the protocol lets you delegate it to a "facilitator" service. The famous one is Coinbase's — which demanded business verification that turned out to be unavailable in Australia. The rescue: the ecosystem has public, keyless facilitators. We point at one URL, POST it the payment, it does the cryptography. No account, no API key, no onboarding call.
A store agents can't find is a hard drive. We shipped a public /agents page ("yes, your agent can buy from us"), a machine-readable manifest at a standard well-known URL, a "For AI agents" link in every footer, and updated our terms for instant-delivery digital goods. Agent traffic today mostly comes from directories and protocol indexes, not Google — being listed is the game.
We promised honestly, so: three things fought back.
One. The spec and the live infrastructure disagreed. The reference implementation described one shape for the message you send a facilitator; the live facilitators wanted another. Our first settlement attempts bounced with a cryptic missing parameters until we read the client library's source and found the field had been renamed and restructured. An evening, gone. This is what "the ecosystem is young" actually feels like: the errors are unhelpful and the documentation trails the code.
Two. Geography. Coinbase's business verification — the gateway to their facilitator and their discovery index — isn't available in Australia. Neither was the main testnet faucet we tried. If you're building outside the US, assume any step involving a US platform's KYC may dead-end, and design around it. The keyless facilitator wasn't our plan A; it turned out to be the better architecture anyway — one less vendor account to maintain.
Three. Tax is not optional just because the money is novel. USDC revenue is still revenue to the tax office. GST treatment of digital goods sold to (probably offshore) software is genuinely murky. We asked our accountant before going to mainnet and the company wallet is in the books from day one. Do this step. It's one email.
What it cost: about 350 lines of code, $0/month in infrastructure (the functions ride free on hosting we already pay for), no new vendor accounts, and two working sessions — one to prove the protocol on a test network, one to stock the shelf and go to mainnet.
What it's made: as of publication, nothing. We're not going to pretend otherwise — the store went live this week. Agent-to-business payments are early; the honest expectation is silence for a while, then a weird trickle, then — if the thesis is right — a channel nobody else in our market is even set up to receive.
Why do it anyway? Because the downside is a rounding error and the upside is being first and legible in a channel that's being built right now by the biggest companies in the world. And because we teach this stuff — our Work Lab cohorts now watch an agent buy from our own site, live, in the session on agents. There's no better way to make "software can do commerce now" land than watching money arrive from a robot mid-class.
Four rules we'd hand any operator trying the same build:
Howll — the AI company behind this lab — runs the Work Lab: six weeks, one working automation per week, including the session where the cohort watches an agent buy from this very website. $999 + GST, capped at twelve.