A one-person studio just launched a catalog of 26 pay-per-call APIs for AI agents — and the single most instructive endpoint in it is a hash function. Not a model, not a data feed, not a "smart" anything. A deterministic sha256 wrapped in an x402 payment challenge.
The reason it exists is the thesis of the whole catalog: models generate strings that resemble valid cryptographic hashes but are not, and a wrong hash silently breaks webhook signature verification.
That one sentence captures what the agent economy is about to pay for — determinism, not abundance.
The 26-endpoint experiment
The catalog is _done, from One Scales — a solo builder shipping narrow, single-purpose APIs that machines call, not humans. Every endpoint is $0.01 per call, paid in USDC over x402 on Base or Solana, with no accounts and no API keys: an agent requests an endpoint, receives an HTTP 402 payment challenge, signs a microtransaction from its wallet, and gets the result. That is the entire on-ramp.
Look at what made the cut for those 26 slots:
| Endpoint | Why an agent can't just do it itself |
|---|---|
| Hashing | LLMs hallucinate plausible-but-wrong hashes; webhook verification needs the real bytes. |
| DNS / WHOIS lookup | Deterministic resolution against live infrastructure, not a model's memory of a domain. |
| DMARC validation | Email deliverability and spoofing checks need actual record parsing. |
| OCR | Turning pixels into exact text — a perception task, not a generation task. |
| Screenshots / port scanning / sitemap extraction | Ground truth about a live page or host, fetched now, not reconstructed. |
None of these are "data an LLM could guess." They are all verification against reality: producing the exact bytes, the exact record, the exact text. That is the common thread, and it is why they can charge at all.
The contrast that matters
The market is currently obsessed with catalog size — how many endpoints a marketplace lists. But a catalog with a thousand endpoints where almost none get called teaches less than 26 endpoints each built around a determinism gap. The unit of value is not "one more API"; it is "one more thing the model cannot reliably do itself."
This matches what the payment data already says. Keyrock's May 2026 report found 76% of AI agent payments sit below Visa's $0.30 fixed-fee floor — card rails are structurally incapable of settling a $0.01 hash lookup. Stablecoin micropayments are the only rail that fits, which is why every one of these deterministic endpoints settles in USDC. The pricing floor and the rail choice aren't separate decisions; they're the same decision.
Why we publish our own zeros
The honest version of this story includes the unflattering numbers. Our own marketplace is serving 817,324 total requests and 20,953 trial calls, against 86 on-chain transactions totaling 3.522 USDC. Most of those 817K requests are crawlers, validators, and quote-cachers — the same shelf-placement traffic every x402 endpoint absorbs. The trial volume is real adoption; the settlement volume is the honest state of an early market.
That gap is the signal. It says the scarce resource in machine-to-machine payments is not rails, not wallets, not discovery — it is things worth paying for. And the things worth paying for, as _done's 26 endpoints show, are the ones that close a determinism gap: the hash, the DNS record, the DMARC check, the OCR'd receipt. Build one of those, price it at a penny, and you have a product. List a thousand guesses and you have a museum.
The takeaway: don't optimize for how many endpoints you can surface — optimize for how many things you can make exactly right that a model can't. Determinism is the product. Abundance is the noise.
Cross-posted from minia2a.uk. minia2a is an open marketplace for x402 pay-per-call endpoints — 15 free trial calls per endpoint, 500 credits on registration, self-custody wallet, no KYC. Sources: One Scales _done launch (MarTech Series, Aug 17 2026); Keyrock agent-payments report (May 2026).
Top comments (0)