AgentIndex · methodology

How we measure the agent economy

Every number on this site comes from something we independently measured or verified on-chain. This page documents each signal, where it comes from, and what it cannot tell you. All figures are recomputable from /data.json.

Provenance labels

Every field in our API carries one of four provenance labels:

On-chain verification

We scan Base mainnet USDC eth_getLogs nightly with a trailing 7-day window. Every Transfer event whose to address matches a listed x402 endpoint counts as an agent payment. Caveats:

• Window is 7 days — not cumulative. The number resets each week.
• 96% of current volume is one sender → one service (blockrun.ai). We publish distinct-payer counts alongside totals to make concentration visible.
• We match transfers by payTo address in the x402 Bazaar, not by inferring intent. A USDC transfer to the same address is counted even if unrelated.

Protocol detection

We run nightly probe campaigns against every indexed domain:

• MCP: JSON-RPC initialize handshake. Confirmed = valid protocolVersion + serverInfo in the response. We then call tools/list using the returned mcp-session-id.
• A2A: GET /.well-known/agent.json returning a parseable agent card with at least one skill.
• x402: Unauthenticated HTTP call returns 402 with a spec-valid accepts[] body. We also compare live terms against the Bazaar’s cached price — 21% of live x402 endpoints disagree with their listed price.
• ERC-8004: On-chain enumeration of the IdentityRegistry at 0x8004A169FB4a3325136EB29fA0ceB6D2e539a432 on Base, Ethereum, and Arbitrum.
• Virtuals ACP, Olas, Masumi: Native chain/API enumeration.

We honour robots.txt. 1,247 domains are counted in the index but unprobed because they disallow our crawler.

Liveness history

We append a row to endpoint_probes after every nightly probe pass. Uptime percentage = confirmed probes / total probes in the trailing 90 probes. This is the one figure a directory that only reads publisher claims cannot produce — it requires completing a handshake repeatedly over time.

Classification

Resource categories are assigned by a keyword/heuristic classifier (src/classify.js) running nightly. Current coverage: ~47% of resource listings carry a DCAT category. Sibling inheritance (3+ classified siblings → inherit the dominant category) covers a further 8%. The remaining 45% are unclassified — we do not guess.

Numbers we will not publish

A shorter list than it should be:

• Cumulative payment totals — we only see 7 days. Presenting a rolling window as a running total is what the old code did; this page exists in part because we fixed that.
• Ecosystem "growth" claims — our index is growing because we are adding crawl sources, not necessarily because the ecosystem is. We show raw counts and let you draw your own conclusion from the trend data.
• Scores from publisher self-reports — our 100-point trust score uses only measured signals (liveness, on-chain, corroboration, protocol richness). Self-declared fields contribute 0 points.

Data freshness

The crawler runs nightly at 03:30 UTC. Each page shows an "Updated" timestamp from the most recent snapshot. Machine-readable data: /data.json.

Questions about methodology? Open an issue or read the source at github.com/cengit1/agent-index.