10,000 AI agents registered on Ethereum. 19 show signs of real activity.

Crypto's pitch for AI agents goes beyond payments: agents should have identities other machines can check, reputations built from past work, and listed services. ERC-8004, a proposed Ethereum standard titled 'Trustless Agents,' provides those pieces: each agent gets an NFT-style identity token, linked to metadata and to a registry where counterparties leave feedback.

The study

'From Agent Identity to Agent Economy' (arXiv:2606.12128, June 10, 2026), by Rischan Mafrur of Western Sydney University and Priagung Khusumanegara of Telkomsel, analyzes the first 10,000 ERC-8004 agents registered on Ethereum mainnet, using on-chain data from January 29 to April 9, 2026. It asks a simple question: beyond registering, do these agents show evidence of doing anything?

The readiness funnel

The authors score each agent on five observable signals: metadata, declared services, reputation feedback, cross-chain registration and transfers. The results narrow quickly.

Their summary: adoption is 'registration-heavy but operationally shallow.'

Concentrated at every layer

Ownership: the 10,000 agents belong to 394 wallets. The median wallet holds 3 agents, the mean is 25.38, and the largest holds 779. The top 10 owner wallets hold 51.40% of all agents, and the ownership Gini coefficient is 0.863 nContext: n=10,000 ERC-8004 agents · Jan 29 to Apr 9, 2026 · arXiv:2606.12128 · margin: not disclosed.

Reputation: there are 980 feedback records from 197 client addresses. The single largest client wrote 645 of them, or 65.82%, and the top five clients wrote 92.4% nContext: n=980 feedback records · Jan 29 to Apr 9, 2026 · arXiv:2606.12128, Table 2 · margin: not disclosed. Only 18 agents received feedback from more than one client.

Services: among agents that declare services, plain web endpoints make up 56.3% of service records, while agent-native formats are rare: agent-to-agent (A2A) 8.0%, MCP 2.7%, x402 payments 1.8% nContext: n: service records of the 67 agents with services, total not stated · Jan to Apr 2026 · arXiv:2606.12128, Figure 8 · margin: not disclosed.

What predicts feedback

The authors also ran an exploratory regression on which agent features go with receiving feedback. The strongest positive associations were linked lifecycle metadata and cross-chain registration; counts of services and service types had smaller positive associations. The simple yes-or-no indicators for having metadata or services came out negative, which the authors attribute to sparse, overlapping data rather than any real effect. They stress the results are associations, not causes: visible, promoted agents may simply be more likely both to publish metadata and to attract reviews.

Limits the authors flag

The dataset is an early snapshot, from the first ten weeks of mainnet registrations. Only a small subset of agents had metadata the authors could fully parse, and feedback may reflect developer testing or one platform's review process rather than market-wide trust. They present the paper as a baseline for tracking the standard as it matures.

A broader census

A second study, 'The Web4 Agent Economy' (arXiv:2606.25876, revised September 2, 2026), by researchers including Jiachi Chen, analyzed 99,448 multi-chain identity registrations and 317,596,323 transaction logs across agent identity, payment and tool protocols. It concludes that machine-to-machine payments are 'highly active' but rest on immature infrastructure, with payment interoperability the most persistent unresolved problem in developer issue trackers.

Reading the two together

A registration is a minted token, not a working service, and both studies predate any large-scale test of agent reputation in markets. The Ethereum data shows how few registered agents expose a service, earn independent feedback or connect across chains, and how much of the activity traces back to a handful of wallets and reviewers.