WorkHeadlinerApr 2026
x402 Verifier Facilitator
x402 facilitator prototype with an LLM-as-judge work-quality endpoint. Four testnet settlements; the verdict never got wired into /settle.
The verifier dashboard: browser wallet on testnet, agent USDC balance, and the live payment path from demo service through the local facilitator to settlement
- The verifier dashboard on a phone
01What I built
Every x402 facilitator shipping in spring 2026 verified the payment: signature, nonce, settlement. None verified the work. For subjective agent tasks like writing or analysis, a signature-verified settlement is not enough; the agent paying needs a quality gate. So I set out to build a third-party facilitator that proxies the reference x402.org facilitator and adds an LLM-as-judge step before /settle. What actually shipped is narrower, and the README says so: the judge runs as a separate /judge endpoint the seller calls, while my facilitator's /verify and /settle are still pass-through proxies. V1 proved the judge call and the settlement path separately; wiring the verdict into /settle was the next step and didn't ship.
What shipped between 2026-04-20 and 2026-04-29: a pnpm monorepo with a facilitator, a paid demo service, a buyer agent and a browser dashboard. Four testnet settlements, all with transaction hashes in the README: a baseline x402 payment on 2026-04-23, the same payment routed through my facilitator-in-the-middle, and two dashboard-triggered agent runs. Seven task-type endpoints were scoped (structured extraction, code, summarisation, translation, creative, strategic, generic). The judge has one rubric of the seven, and it is not yet in the settlement path. It stalled on 2026-04-29 and I have written it up that way rather than as a finished product.
02What I learned
Three things changed my March assumptions. I always believed Solana would win, and it happened faster than I planned for: Solana passed Base on daily x402 volume in January 2026. The lesson was to build for any chain from day one. The category where verification matters most, AI-generated services, had almost no volume ($14.2k against $548.5k in agent-to-agent services). And the judge does not pay for itself at micropayment scale: $0.01–$0.15 per evaluation against a $0.20 average transaction is 5–75% of the value. Whether agent transactions grow past about $5 is the single load-bearing assumption in the whole thesis.
I also learned what a real prototype costs in operator time: funding testnet wallets, chasing faucets, reading facilitator responses. Doing it myself is why I can now explain the settlement path to a partner in plain language.
03Built with AI
Built entirely by directing Claude Code against a twelve-section implementation plan written in the research vault first. I never read the code. I judged it by whether the transactions settled on the testnet, whether the dashboard showed the payment path live in a browser, and whether the README matched what I saw. The four transaction hashes are the evidence.
Solana relevance
The design is chain-abstract because Solana was already leading x402 volume by the time I built it. The gap it targets, outcome verification on top of a payment standard, is the one piece missing from Solana's agentic-payments stack as of September 2026: the network has settlement, attestation-based identity and roughly 65% of x402 volume, and nothing that checks an agent did what it was paid for. Pay.sh meters access per call; that is not the same thing.