Identity Expansion: Participation Depth in ERC-8004

1 · Introduction

If AI agents are to become economic actors, they must first be identifiable, discoverable, payable, reviewable, and trustworthy. ERC-8004 (Trustless Agents) is the Ethereum standard created for this purpose [1]: each agent receives an ERC-721-style on-chain identity token, connected through Identity, Reputation, and Validation registries to metadata, service declarations, and historical feedback. After its official mainnet launch on January 29, 2026, registrations passed ten thousand within roughly ten weeks.

The existence of registrations, however, is not the same as a functioning economy: a registered agent can be an empty shell. To support open research on this question, this lab compiled and released a complete dataset of the first 10,000 ERC-8004 agents registered on Ethereum mainnet—eight tables covering identity records, mint economics, transfer history, reputation summaries, and feedback details—published as an arXiv preprint with a Harvard Dataverse archive [2, 3]. This report presents the first independent measurement results built on that dataset [4] and interprets them through the series' layered authenticity framework.

2 · Measurement Framework: Layered Observable Evidence—and What It Does Not Measure

Mafrur and Khusumanegara [4] construct on this dataset what they term an “operational readiness” framework: identity registration is the baseline, and an agent exhibits observable evidence of ecosystem participation layer by layer as it provides parseable metadata, declares service endpoints, accumulates reputation feedback, or appears in cross-chain registration and transfer records. The boundary of this framework must be stated clearly: it measures observable on-chain participation depth—not whether a service is actually live, whether endpoints respond, whether tasks are completed, whether payments are received, and certainly not whether decisions are made autonomously by AI. The original study itself positions these as descriptive, exploratory proxy indicators [4]. The data are moreover a cross-sectional snapshot (blocks 24,339,925 to 24,839,925, i.e., January 29 to April 9, 2026; agent IDs 0–9999) and do not capture dynamics after the window [3, 4].

3 · Core Finding: Registration-Heavy, Observably Shallow

The measurements form a steep funnel: of 10,000 registered identities, only 628 (6.28%) have received at least one piece of reputation feedback; only 67 declare at least one service record; and only 19 combine all four kinds of observable evidence [4]. The overwhelming majority of registered identities remain at the top of the funnel—in the framework of this series, at L1.

Concentration metrics further undercut the appearance of broad adoption: the 10,000 agents map to only 394 owner wallets; the top 10 wallets hold 51.40%, the largest single wallet holds 779, and the ownership Gini coefficient is 0.863. Concentration in the reputation layer is even more extreme: all 980 feedback records come from 197 client addresses, the largest single client contributes 645 (65.82%), the top five together account for 92.4%, the feedback Gini coefficient is 0.783, and only 18 agents have feedback from more than one client [4].

Concentration by itself, however, does not identify a cause. Top-10 wallets holding over half of all agents is consistent with manipulative batch minting—but equally with a single developer registering many distinct agents, a marketplace registering on behalf of its users, test addresses, or a legitimate fleet operated by one company. These mechanisms have very different economic implications, and separating them requires deployer attribution and longitudinal activity analysis, which a cross-sectional snapshot cannot provide. Likewise, extreme feedback-client concentration matches the signature of Sybil manipulation but also the early-ecosystem reality that a single platform may currently be the only active reviewer.

The composition of the service layer offers another clue: among the minority of agents declaring endpoints, plain WEB endpoints account for 56.3%, while the protocols actually designed for machine-to-machine collaboration—A2A (8.0%), MCP (2.7%), and x402 (1.8%)—remain a small minority in aggregate; endpoint domains concentrate heavily in a few platforms such as marketplace.olas.network [4].

4 · Interpretation: The Identity Layer Is Not the Economic Layer

The evidence supports a restrained conclusion: the mainnet deployment of ERC-8004 currently functions more like an on-chain identity registry than operating agent-economy infrastructure. This does not diminish the standard's value—identity is the precondition for any subsequent trust mechanism—but the transition from identity registration to broad, decentralized, and verifiable economic activity remains at a very early stage [4]. Cross-sectional on-chain data also have inherent limits—endpoints may go undeclared in registration files, and interactions may occur off-chain or after the data window—so what the evidence supports is that a broadly functioning agent economy is not yet visible, not that no economic activity exists. Consistent with this, an independent empirical study by Xiong et al. covering Ethereum, BNB Smart Chain, and Base (data through May 13, 2026) reports sharper findings still: only 3%, 4%, and 15% of registrations on the three chains, respectively, expose a valid registration file with at least one live service endpoint; 73.5%, 59.2%, and 90.6% of reviewers, respectively, exhibit coordinated Sybil behavior; no mainnet deployment of the Validation Registry was observed within their window; and the Reputation Registry, as currently deployed, cannot function as a trust signal because reputation can be manipulated at minimal cost [5].

Placed in the layered framework: ten thousand registrations are L1 evidence; feedback records and endpoint declarations provide participation evidence beyond bare registration, but they neither directly establish independent economic activity (L2) nor service delivery (L3); and the autonomy dimension (L4)—who or what controls these identities and interactions—lies entirely outside what these data can measure. What deserves tracking is therefore not the headline registration count but movement in the middle and lower funnel layers, the diversity of feedback sources, the share of agent-native protocol endpoints, and the number of distinct operators after deployer attribution. This lab will continue updating the dataset to provide a longitudinal measurement baseline for that transition.

Data & methods notes

• On-chain data: Ethereum mainnet blocks 24,339,925–24,839,925 (January 29 – April 9, 2026), agent IDs 0–9999; a cross-sectional snapshot that does not reflect post-window dynamics.

• Metrics are proxy indicators of observable on-chain evidence; they cannot establish service availability, task completion, or agent autonomy, and concentration metrics do not distinguish batch deployment, marketplace registration, testing, fleet operation, and manipulation.

• Both measurement studies cited here [4, 5] are arXiv preprints (not peer-reviewed).