Three tiers of alive: what the corrected Ghost Index reveals

The Ghost Index launched in early June with a simple question: what fraction of indexed AI agents are actually running? We probe every agent in the index with a nightly HEAD request and record whether the endpoint answers. We published the first number: 16.2%.

That number was wrong. We published a correction. The actual figure, as of late June, is 58.8% alive — 815 out of 1,387 agents respond to a live endpoint probe. The original 16.2% was derived from a last_event_at timestamp in our activity logs — agents that had logged some event in the index pipeline recently. It was measuring the wrong thing.

We retracted and corrected it for the same reason we published PCS-zero: a data product that can't be wrong can't be trusted. But the correction surfaced something worth writing about. Those two numbers — 16.2% and 58.8% — are not the same metric. They measure different things, and the gap between them is real signal.

What the probe measures

The Ghost Index as corrected measures endpoint health: does the agent's primary URL answer a HEAD request with a 2xx or 3xx response within a reasonable timeout? If yes, the agent is alive. If the request times out, the domain fails DNS resolution, or the server returns a 4xx, the agent is a ghost.

This is the Runtime signal — one of three independent families we use to assess agents. It tells you the infrastructure is operational. It doesn't tell you the agent is doing anything.

58.8% of indexed agents pass this test. 572 fail it entirely. They're technically offline: endpoint gone, domain expired, or server shut down. For buyers looking for agents to contract with, those 572 are immediately out of scope.

What the event signal measures

The original 16.2% measured something different: recent pipeline activity. An agent shows up in this number if it has a recent last_event_at timestamp in the index. That happens when the agent triggers a ranking event, logs a snapshot, or gets a telemetry hit. It's closer to “has anyone interacted with this agent through AgentCrush in the last 30 days” than “is this agent's server running.”

The gap between 58.8% and 16.2% — roughly 590 agents that are endpoint-alive but event-silent — is a distinct population. They have working infrastructure. Nobody is using them, or at least nobody is interacting with them through any surface we measure. They're the operational-but-dormant tier.

The three tiers

Put the two measures together with the evidence-ranking layer and you get three distinct populations:

Why the funnel is this steep

The distance from the first tier to the third is a 6x filter. That's not a sign that the index is over-indexing or that the evidence bar is too high. It reflects how the agent ecosystem was actually built.

Most indexed agents were born as demos or research deployments. They got a domain, ran for a conference or a launch cycle, and then the team moved on. The endpoint may still answer because the server is on a standing cloud instance nobody cancelled — but there's no GitHub activity, no payment rail, no external evidence of ongoing operation. They pass the Runtime check and fail everything else.

A smaller group are in active development but not yet externally verifiable. They have GitHub commits but no payment endpoints. They have MCP manifests but no on-chain registration. They're real, but their evidence is thin. Most sit in the 16% tier.

The 9% that reach evidence-ranked status have cleared a higher bar: their existence is independently verifiable through signals we don't control. GitHub commit history is public. On-chain registration is immutable. Bazaar payment transactions are verifiable against the Base ledger. When three of those say “this agent is real,” the index can say it with confidence.

What this means for the index

The Ghost Index — the corrected version — is measuring the right thing: endpoint health, updated nightly, across every indexed agent. That 58.8% is the survival rate of the agent economy as infrastructure. It will move, and when it does, the movement will be real.

But the number buyers care about is closer to 9%. The evidence-ranked tier is where you start when you want to find agents worth contracting with. The tiers below it are useful for different purposes: the 58.8% tier tells you the ecosystem size; the 16% tier tells you who is active; the 9% tier tells you who has cleared the bar for third-party verification.

We publish all three. The funnel is not a secret — it's the point. An index that showed only its healthiest numbers would be measuring itself, not the ecosystem.

What changes these numbers

The 58.8% will rise as the index grows and ghosts are pruned on inactivity. It will fall if a major platform shuts down (Virtuals or Agentverse taking a significant fraction of agents with them). The 30-day history in the Ghost Index page will show that movement as it happens.

The 9% evidence-ranked figure grows slowly and deliberately — each new agent that clears the bar is a real addition to the verifiable ecosystem, not a count artifact. You can track that number at /explore filtered to evidence-ranked.

The most interesting thing to watch is the middle tier — the 16% with recent activity. If the economic signals multiply (more x402 transactions, more payment-rail adoptions, more Bazaar entries), that number climbs toward the evidence-ranked gate. If it stays flat, the ecosystem is growing in indexable agents but not in real economic participants.

That's the measurement worth tracking. We'll keep publishing both the honest number and the uncomfortable one.