Ask three different people in the AI agent ecosystem how many AI agents currently exist, and you'll get three different answers — because they're looking at three different registries. ERC-8004 reports 197,000+. ClawHub's marketplace has roughly 8,200 listed agents. Virtuals shows 2,300+ tokenized agents on Base. These numbers don't add up to a coherent picture. They add up to the fragmentation problem.

Three Registries, Three Different Worlds

ERC-8004
On-chain · Permissionless
197K+
Any agent can register. No curation. Metadata quality varies widely. Ground truth for on-chain identity.
ClawHub
Off-chain · Curated marketplace
~8,200
Quality-filtered, reviewed listings. Richer metadata (pricing, reviews, demos). Primarily skill/tool agents.
Virtuals
On-chain · Tokenized
~2,300
Tokenized agents with market caps. Base network. Rich financial metadata but narrower scope.

Each registry is optimizing for something different. ERC-8004 maximizes permissionless discoverability — any agent can register without gatekeepers. ClawHub optimizes for marketplace utility — humans can find and compare agents quickly. Virtuals optimizes for tokenized ownership — agents as investable assets with on-chain economics.

None of these goals are wrong. The problem is that they've created three incompatible namespaces for the same underlying entity: an AI agent.

The Same Agent, Three Different Identities

Here's the concrete problem. Take a mid-tier trading bot, "TradingBot v3," deployed by a solo developer in February 2026:

In this case, the same agent has two separate registry entries and no obvious link between them. The ERC-8004 address and the ClawHub listing don't reference each other. A researcher looking at ERC-8004 counts "0x4f2a..." as one agent. A researcher looking at ClawHub counts "mkt-bot" as another agent. They're the same thing, counted twice in any aggregate analysis.

The double-counting problem at scale: Our current estimate is that 12,000–18,000 agents appear in more than one registry. Without cross-registry matching, every aggregate count of "total agents" overcounts by at least that amount — and the true overlap is probably larger.

Why This Matters Beyond Statistics

Double-counting is annoying but fixable with better methodology. The deeper problem is what fragmentation prevents:

Due diligence without a unified profile is guesswork

If you're a protocol considering integrating an agent, or an investor considering funding an agent team, you want to know: Is this agent actually active on-chain? Does it have a verifiable track record? Is its claimed capability description consistent across platforms? You can't answer these questions by looking at a single registry. You have to manually correlate data across sources — which most people don't do.

Competitive intelligence is distorted

A trading agent that appears on ClawHub but not ERC-8004 looks less serious than one that appears on both — even if they're functionally identical. Market share analyses, capability distribution studies, chain adoption metrics — all of these are skewed by the uneven cross-registry presence of different agent types.

Agent reputation can't be portable

When an agent's identity is fragmented across registries, its reputation can't travel with it. A 4.8-star rating on ClawHub can't be verified or referenced from an ERC-8004 smart contract call. An agent with a strong on-chain transaction history can't surface that history in its ClawHub listing. Reputation becomes silo'd, which means agents have to rebuild trust separately in every ecosystem they enter.

What a Unified Index Would Change

Cross-registry matching — connecting the same agent's entries across ERC-8004, ClawHub, Virtuals, and future registries — changes the picture fundamentally. Instead of three separate counts, you get one canonical profile per agent that aggregates signals from all sources.

This is what Trustprint is building. The matching problem is technically hard: ERC-8004 addresses don't carry the same identifier as ClawHub slugs. You need multi-signal matching — wallet address correlation, metadata similarity, operator identity verification, behavioral pattern matching — to connect entries that represent the same agent across different namespaces.

The result, when it works, is the first time you can ask meaningful questions like: "How many unique, active AI agents exist across all major registries?" or "Which agents have the strongest cross-registry presence?" or "What's the overlap between Virtuals tokenized agents and ERC-8004 registered agents?"

Right now, no one can answer those questions with confidence. The three-registry problem makes the true scope of the agent economy genuinely unknowable — until someone builds the layer that connects them.