# Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem

Crypto.world

> Autonomous AI agents increasingly need to transact with strangers, but they have no reliable way to judge whether an unknown counterpart is trustworthy. ERC-8004, deployed on Ethereum mainnet on 29…

**arXiv**: 2606.26028

**Published**: 2026-06-24

**Authors**: Xihan Xiong, Zelin Li, Wei Wei, Qin Wang, William Knottenbelt, Zhipeng Wang

## AI summary

## Overview

**Research area:** Blockchain systems security and empirical measurement, specifically decentralized trust and reputation infrastructure for autonomous AI agents (cs.CR), with ties to AI agent economies, ERC-8004, and x402 payments.

**Technical level:** Intermediate. Readers need basic familiarity with Ethereum-style smart contracts, NFTs, and gas costs, but the paper's core argument is accessible without deep cryptographic background.

**Scope (one sentence):** The first cross-chain empirical measurement of every Identity and Reputation Registry event in the ERC-8004 "Trustless Agents" protocol on Ethereum, BNB Smart Chain, and Base, from deployment through 13 May 2026, assessing whether registered identities are real and whether recorded reputation can serve as a trustworthy signal.

## What This Paper Is About

Autonomous AI agents increasingly need to transact with strangers, but they have no reliable way to judge whether an unknown counterpart is trustworthy. ERC-8004, deployed on Ethereum mainnet on 29 January 2026, attempts to solve this with three on-chain registries for Identity, Reputation, and Validation, and it attracted more than 170k registered agents and over 150k feedback records within months. The paper asks whether this fast-growing on-chain data actually means anything: who registers agents, whether those identities correspond to working agents, and whether the reputation records can be trusted by a decision-making agent.

## Key Contributions

1. **A multi-chain dataset.** The authors assemble the first comprehensive ERC-8004 dataset covering every Identity and Reputation event on Ethereum, BSC, and Base, enriched with off-chain registration and feedback files, per-transaction gas costs, and x402 settlement transactions. Researchers may request access for reproducibility.
2. **Identity and adoption analysis.** They characterize registration dynamics, batch minting, ownership concentration (Gini 0.733 on Ethereum, 0.708 on Base, 0.133 on BSC), the gap between minting and activation, and the quality of registration files and declared services across the three chains.
3. **Reputation analysis, including a security evaluation.** They test the Reputation Registry against four necessary conditions for a trust signal (commensurability, robustness, groundedness, economic soundness), measure the cost of fabricating or destroying reputation, and quantify Sybil behavior in the live data.
4. **Actionable recommendations and responsible disclosure.** Each finding is converted into a concrete design change for the next ERC-8004 revision, and the findings were disclosed to the ERC-8004 proposer (the Ethereum Team), with their feedback incorporated.

## Main Findings

- **Registration volume is not ecosystem maturity.** Cumulative registrations reach 32,343 on Ethereum, 90,145 on BSC, and 50,985 on Base (173,473 total), but only 3%, 4%, and 15% of agents respectively expose a valid ERC-8004 registration file with at least one live service endpoint.
- **Compliance and functionality diverge by chain.** Only 29.3% of Ethereum agents, 83.3% of BSC agents, and 26.9% of Base agents have a valid ERC-8004 registration file. BSC achieves high file compliance (79% valid) yet only 4% declare a service endpoint; Base has the most service-enabled agents (15%) but also the largest share of non-compliant files (25%).
- **Batch minting dominates Ethereum registration.** 446 batch transactions (2.6% of registration transactions) created 15,607 agents, or 48.3% of all Ethereum agents, with a maximum batch size of 182 and median of 10. BSC shows the opposite pattern: 11 batch transactions registering 22 agents. Base sits between, with 829 batch transactions creating 3,966 agents (7.8%).
- **Ownership is highly concentrated on Ethereum and Base.** The top 1% of wallets own 58.5% of Ethereum agents and 54.9% of Base agents, and the top 10% control more than 70% on both. BSC's top 1% and top 10% own just 12.1% and 22.1%.
- **Platform-mediated registration explains concentration.** Nookplot received 6,975 identities (transferring 6,909 of them to distinct addresses); UniAgent holds 1,243 asset-linked identities; Olas holds 605; Execution Market transferred 125 of 194 NFTs to 40 addresses. Seven platform-associated cases span eight addresses and 9,726 agents (20.15% of the top-50 sample, 5.61% of all 173,473 registrations).
- **Most identities are never activated into agents.** 53% of Ethereum agents and 37% of Base agents were never activated (empty URI), versus 9% on BSC, where 91% were active at mint. Post-mint activation accounts for only 1% to 13% of registrations, and when it happens it is fast: 92% of Ethereum and 93% of BSC activations occur within one day, while on Base 41% update in the same block and another 43% within a day.
- **URI schemes differ sharply.** BSC is 81% inline `data` URIs; Ethereum and Base use roughly 29% `data` URIs, and Base is the only chain with meaningful IPFS adoption (5%) alongside 27% web-hosted files.
- **Advertised functions are narrow and templated.** Of 98,702 agents (56.9%) with usable descriptions, only 14,530 distinct texts remain and the top 50 templates cover 77.3% of agents. Trading & DeFi accounts for 67.1% overall, with two templates covering 85.2% of that population on BSC; 91.2% of BSC agents with usable descriptions advertise trading functions, while 49.3% of Ethereum descriptions are generic or vague and 24.2% are non-agent text. Base is the most diverse, hosting analytics, developer tools, security, and agent infrastructure.
- **The Reputation Registry meets none of the four necessary conditions for a trust signal.** Values are not commensurable, the aggregate can be moved by a single input, feedback is rarely tied to verifiable interactions, and reputation can be fabricated or destroyed at minimal cost: a median of $0.055, $0.0042, and $0.0027 on Ethereum, BSC, and Base respectively.
- **Sybil reviewing is pervasive.** Sybil-flagged reviewers make up 73.5%, 59.2%, and 90.6% of reviewers on Ethereum, BSC, and Base, affecting the displayed reputations of 26.4%, 81.4%, and 96.2% of rated agents.
- **Little reputation survives Sybil removal.** After removing Sybil-flagged feedback, 15.8%, 77.9%, and 86.8% of rated agents on Ethereum, BSC, and Base are left with no valid feedback.
- **The reputation market is unevenly distributed.** Base dominates with 122,798 feedback records from 3,073 reviewers across 28,592 rated agents, versus 29,444 records from 76 reviewers across 4,310 agents on BSC and 3,058 records from 618 reviewers across 1,559 agents on Ethereum. BSC reviewers are extreme outliers, averaging 387.4 feedback records and 56.7 agents each, compared with 40 and 9.3 on Base and 4.9 and 2.5 on Ethereum.
- **Feedback distributions and vocabularies are inconsistent.** Ethereum spikes at a score of 100; BSC concentrates between 70 and 80; Base is the most dispersed with mass at both 100 and near zero, and is the only chain with out-of-range values (n=1 below zero, n=2 above 100). Ethereum's tags are the most diverse, BSC's are repetitive and placeholder-like, and Base's center on risk vocabulary such as trustscore and contractrisk.
- **Reputation does not travel across chains.** Comparing agents registered on multiple chains reveals no significant cross-chain correlation, indicating fragmented reputation silos.
- **The Validation Registry is not yet observable.** No confirmed mainnet deployment existed during the observation period, so the high-assurance tier of the protocol could not be measured.

## Methodology in Plain English

The authors treat the protocol as a natural experiment and measure what actually landed on-chain. They pick the three chains that hosted ERC-8004 with the highest activity, set a common end time (13 May 2026, 21:49:47 UTC), and choose start blocks before the earliest deployment so no events are missed. From the Identity Registry they collect `Registered`, ERC-721 `Transfer` (mint and secondary), `URIUpdated`, and `MetadataSet` events; from the Reputation Registry they collect `NewFeedback`, `FeedbackRevoked`, and `ResponseAppended`. Replaying URI updates in block order reconstructs each agent's current URI, and replaying feedback events reconstructs each agent's feedback set.

They then follow every pointer off-chain, downloading and parsing the referenced registration and feedback files to check whether files are retrievable, whether the `type` field matches the canonical ERC-8004 identifier, and whether any service endpoints are declared. To compare manipulation costs across chains, they compute gas for five event types and convert native-token costs to USD using hourly ETH/USDT and BNB/USDT closing prices from Binance, matched by nearest candle within a two-hour tolerance. Settlement data comes from x402 payments attributed to agents via their designated payout wallet.

For the reputation security analysis, they define what a trustworthy score would require, then test the deployed contract against each condition and measure real-world manipulation by flagging coordinated Sybil reviewers and recomputing what reputations look like without them. For the functional analysis of agent descriptions, they deduplicate near-identical text (cosine similarity at least 0.95), embed with the bge-m3 model, reduce with UMAP, cluster with HDBSCAN, label clusters into 12 categories, and propagate labels to the long tail with a trained classifier.

## Why This Matters

**Impact on research.** The paper establishes an empirical baseline for studying AI agent markets and shows that on-chain registration counts, the most visible adoption metric, are a poor proxy for real deployment. It also provides a reusable measurement template: define necessary conditions for a trust signal, then test a deployed protocol against them rather than against its specification.

**Real-world applications:**

- Agent marketplaces and orchestrators can use the quality taxonomy (no URI, unresolvable, non-compliant, valid without service, valid with service) to filter counterparties before committing funds.
- Reputation aggregators building on ERC-8004 can adopt client-set and tag filtering as mandatory rather than optional, given how much displayed reputation disappears after Sybil removal.
- Protocol designers in adjacent standards can reuse the four-condition framework (commensurability, robustness, groundedness, economic soundness) to stress-test reputation designs before deployment.
- Enterprises deploying agent-to-agent payments over x402 can use the cost figures to reason about how cheaply an adversary can manufacture a favorable score.

**Industry relevance.** ERC-8004 is positioned between communication protocols (A2A, MCP) and payment rails (x402), making it infrastructure that other companies will build on. The finding that manipulation costs fractions of a cent, and that most reviewers on Base exhibit Sybil behavior, is directly relevant to anyone pricing risk or writing procurement rules around autonomous agent transactions.

## Future Directions

- **Revise the Reputation Registry.** The paper offers concrete implementable changes for the next ERC-8004 revision, though the specific recommendations fall in the section beyond the excerpt provided here; the identified gaps include missing reviewer-reputation infrastructure, non-standardized tag semantics, and incompatible scales under identical tags.
- **Measure the Validation Registry once deployed.** The high-assurance tier (stake-secured re-execution, zkML proofs, TEE attestation) had no confirmed mainnet deployment during the study, leaving open whether validation evidence changes agent selection behavior.
- **Build defensible Sybil resistance and reviewer reputation.** The specification delegates Sybil resistance and aggregation to off-chain systems, but no such reviewer-reputation layer yet exists, and reviewers remain unconstrained apart from the ban on self-rating.
- **Bridge fragmented cross-chain reputation.** The absence of significant cross-chain correlation among agents registered on multiple chains raises the question of whether portable reputation can be constructed, and how self-declared `registrations` links could be verified.

## Target Audience

Protocol designers and security researchers working on decentralized identity and reputation; AI infrastructure engineers building agent marketplaces on A2A, MCP, or x402; blockchain data analysts interested in measurement methodology; and standards contributors who need evidence on how trust primitives behave once deployed rather than as specified. Readers evaluating whether to rely on ERC-8004 reputation today will find the most directly actionable material.

## Authors’ abstract

As autonomous AI agents increasingly transact across organizational boundaries, a fundamental trust challenge emerges: how can an agent assess whether an unknown counterpart is trustworthy? The ERC-8004 protocol addresses this challenge with the first permissionless trust layer for AI agent economies, built around three on-chain registries for Identity, Reputation, and Validation. Despite its rapid adoption, the protocol has not been studied empirically, leaving it unclear…

[The full abstract on arXiv](https://arxiv.org/abs/2606.26028)

[Read the original paper](https://arxiv.org/abs/2606.26028)

---

Canonical: https://crypto.world/paper/can-trustless-agents-be-trusted-an-empirical-study-of-the-erc-8004-2606.26028
