Agent Economy with receipts

Trust & Safety · How to evaluate a seller

How do I check a seller before my agent pays it?

With public evidence, one question at a time. Every check on this page can be run today, for free, with a browser — no account, no tools to buy. Each check is a question, and each names the public evidence that answers it.

Where do these checks work best?

They work best where payments are public. As of 2026-09-02, x402 payments settle in USDC on the Base network, so a seller’s payment history sits on-chain — on a public ledger anyone can inspect. Card-rail purchases leave no public ledger; there, the rail’s built-in dispute process is the backstop instead.

One limit to carry through the whole page: public payment history proves that money moved. It cannot prove the work was good — that is the verification hole. These checks lower the odds of paying a fake. They do not grade anyone’s work.

Check 1: Does the seller publish a payment address?

An on-chain seller has a receiving address, and a checkable seller does not hide it. Look in the seller’s catalog listing and its documentation. If your agent has already paid once, the settled transaction names the recipient — the address is on your own receipt.

Check 2: Does on-chain history show buyers who come back?

A working service accumulates history that is hard to fake well: many distinct buyers, arriving over months, paying amounts that match the published price — and returning. Paste the seller’s address into a public Base block explorer, such as BaseScan, and read the incoming transfers.

Reading a ledger takes patience, and that is fine. The point is that you can — and that no seller can stop you.

Check 3: Is the pricing sane?

A price has to make sense twice: against comparable services, and against itself. First compare the seller’s published price with the published prices of comparable services. Then confirm that the price its service actually quotes matches its listing — in the x402 flow, the service names its price in the 402 answer it returns when asked.

OTTO, our example AI agent, buys text extraction at $0.01 per page. Its operator runs the self-consistency check in one request: ask the service for its price, and confirm the quote matches the listed $0.01. A seller whose quote and listing disagree has failed a check before any work is judged.

Check 4: Are the seller’s claims sourced?

A claim you cannot check should carry no weight — not treated as false, just as unchecked. Sellers claim request volumes, revenue, uptime, customer counts. For each claim, ask one question: what public record backs this?

After the checks pass, then what?

Spend a little first. Before granting a standing mandate, make one small purchase and read what comes back. The receipt cannot grade the work — but you can.

Then cap it. Keep a spending cap on the agent no matter how good the checks looked. The checks lower the risk; the cap bounds it.

This checklist is manual today. The same signals — buyer spread, dust filtering, price sanity — are what this site’s published numbers will be built from. The method is public in the methodology.

Where are the primary sources?

What changed on this page?