The traffic tripwire observed

A neutral verifier's market switches on when agents transact across trust boundaries — the moment a referee becomes necessary, not just nice. That "when" is the biggest strategic unknown, so we measure it. These are the leading indicators, re-read monthly; an alert fires the moment any crosses its threshold.

What this reads today (2026-08). About 1 in 10 live MCP server domains also publishes a valid A2A agent card (unbiased sample) — far more than A2A's own advertised partners, where cards barely exist. (An earlier reading of 16.5% was inflated by alphabetical sampling and corrected to 9.5%.) That is cross-protocol addressability, not traffic. The layers that turn addressability into traffic — a real discovery directory, a standardised identity scheme — are still flat. Substrate appearing; cross-org traffic not yet. The honest reading: not yet time to lean in — you'll be told when it changes. One more thing is now watched here: the paid tool economy. It has appeared but not diversified — 13 operators behaviorally demand payment today; the wire trips at 25.

A2A agent-card density quiet

0%trips at 25%40%

10.0% — 20 of 200 live MCP domains

Of live MCP server domains, how many also publish an A2A agent card — cross-protocol addressability. A leading indicator, not traffic itself. That is ~16× the 0.6% found across A2A's own advertised partners: cross-protocol readiness is real, but a published card is a doorbell installed, not a visitor arriving.

A public A2A directory quiet

today

unreachable

trips

a real, populated directory exists

Discovery: is there a real, populated A2A agent directory, or still a parked domain? Traffic needs somewhere to discover peers.

Agent-identity standardisation quiet

today

0 adopted

trips

the first IETF draft is adopted

Adopted IETF Web Bot Auth drafts — the only shipped agent-identity scheme moving from proposal to standard.

Paid tool economy — operators quiet

0trips at 2530

13 — 13 operators

Distinct operators (registrable domains) whose servers behaviorally demand payment on a tool call — strict evidence: HTTP 402 or x402 payment fields in the response body, never a name, never a header. Money on tool calls is what creates the first party with standing to dispute a wrong answer; verification demand follows liability. Those operators run 114 paying endpoints between them — the gap between the two numbers is the costume-farm pattern at the payment layer (on day zero one operator was 95 of 99 endpoints). Method pre-registered; strict classifier; a name marker or an advertised price never counts — only a demand.

For scale: 10,813 remote MCP endpoints exist in all — context, not a tripwire; it only grows.

First honest reading · 2026-08. We show no trend line yet: a reading needs ≥2 genuine monthly captures, and the earlier 16.5%→9.5% move was a sampling correction, not a change in the world — plotting it would assert a decline that did not happen.

Cards are supply-side readiness, not transactions: a published agent card means an agent can be reached peer-to-peer, not that anyone is transacting across orgs. We never conflate the two.