Crawlers vs. wallets: what our first day of payment data says about the agent economy

Yesterday at midday we switched on telemetry at our payment gate. Every AgentCrush API response now increments a counter: free calls, price quotes on the paid endpoints, settled x402 payments, subscriber calls. By midnight the gate had served 1,376 price quotes to machines. Settled payments: zero.

Most companies would not publish that second number. We think it's the most interesting thing we've measured all month.

Who asks for 1,376 prices and buys nothing?

The access pattern gives it away: almost all of the traffic replays the exact example URLs we publish in our marketplace listings. These aren't customers browsing. They're the agent economy's infrastructure, doing its rounds:

Marketplace validators. The Coinbase Bazaar re-checks every listed endpoint — is it alive, does the 402 quote parse, did the price change. Those probes are how the marketplace's own quality stats get made.

Catalog builders. Directories index machine-payable endpoints and resell discovery. We know this crawler species well, because we are one — our own index mirrors 46,000+ listed x402 resources.

Security scanners. At least one risk-scoring service swept x402 endpoints this week. We were almost certainly in their sample.

Quote-cachers. Orchestrator agents that record prices now to route purchases later. Window-shopping, machine-speed.

None of them pay. All of them matter: a probe means you exist in the catalogs that buying agents consult. In a machine economy, being crawled is shelf placement. The probes are the prerequisite for purchases, not their substitute.

The wallets are real — they're just elsewhere

The same marketplace data answers the obvious follow-up. Across the listed x402 economy we counted 322,948 paid calls in the last 30 days. One transaction-simulation tool alone had 2,596 unique paying wallets. Agents with funded wallets exist, and they spend daily.

But the demand is brutally concentrated. A handful of providers carry nearly all the real volume — while a single lister accounts for roughly 90% of all listings and almost none of the payers. The agent economy is simultaneously real and mostly noise. Telling those apart is the entire reason AgentCrush exists, so we shipped the obvious thing: an x402 merchant directory ranked by paid calls instead of listings. Ranked by demand, the 32,000-listing host drops to fourth place.

The stat we almost published

One more honesty note, because this is the part nobody writes about. Yesterday we also computed our first Protocol Compatibility Score — how many of the four live agent rails (ERC-8004, x402, A2A, MCP) each indexed agent verifiably supports. First readout: 104 agents at three or more rails. Impressive number. We sampled the list before publishing it and found link-collection repositories "supporting" payment protocols.

Our scanner was counting soft 404s — sites that return a friendly HTML page for any path, including the protocol manifests we probe for. We fixed the scanner to validate response content, re-scanned everything, and the truth is starker than the bug: of 1,359 indexed agents, not one verifiably supports three of the four rails. Exactly one supports two. The "maximally interoperable agent" that every protocol deck assumes? It doesn't exist yet.

Why we publish the zeros

A dashboard that can't show zero can't be trusted at any other value. Our paid-call counter reads zero today. Our interop counter reads zero. Both will move — and when they do, you'll know the movement is real, because you watched us publish the floor.

The numbers in this post are live and free: the daily index diff, the interoperability readout, the Ghost Index. Check our math. That's the point.