ERC-8004 Explained: Trust Standard for Blockchain AI Agents

Introduction

Roughly 15,000 of about 165,000 agents registered under ERC-8004 sat dead or uncallable by April 2026, proof that an on-chain identity is not the same as a working agent. That gap captures both the promise and the immaturity of Ethereum's trust standard for AI agents. ERC-8004 gives autonomous agents three on-chain registries for identity, reputation, and validation that let them discover and vet each other without a central authority. Yet the standard, created in August 2025, remains a draft in 2026, and its validation layer is still unfinished. This guide explains what ERC-8004 does, how its three registries work, where it fits alongside A2A, MCP, and x402, and how finalized the standard really is.

What Is ERC-8004, and What Problem Does It Solve for Blockchain AI Agents?

ERC-8004 is a draft Ethereum standard that lets autonomous AI agents prove who they are, show a track record, and have their work checked without a central authority. Its formal title is EIP-8004: Trustless Agents, and it targets the discovery-and-trust gap in the emerging agent economy.

What ERC-8004 Actually Is

ERC-8004 defines three lightweight on-chain registries that any agent can use across organizational boundaries. Blockchain AI agents already execute trades, hire other agents, and manage portfolios, yet no shared method existed to verify an unknown counterpart. The standard fills that gap with portable identity, reputation signals, and validation records, deployable on any Ethereum Layer 2 or mainnet as per-chain contracts.

Why It Is Still a Draft

Despite active use, ERC-8004 is not a finalized Ethereum standard. It was created on 13 August 2025 as a Standards Track ERC and remains a draft as of mid-2026 (Ethereum Improvement Proposals, 2026). Four co-authors lead it: Marco De Rossi of MetaMask, Davide Crapis of the Ethereum Foundation, Jordan Ellis of Google, and Erik Reppel of Coinbase. Specification details still move rather than sit settled.

Understanding the gap ERC-8004 closes starts with how its three registries fit together.

How Does ERC-8004 Work Through Three On-Chain Registries?

ERC-8004 splits agent trust into three separate contracts—Identity, Reputation, and Validation—so agents discover each other and judge reliability without a central directory. Each registry is a per-chain singleton: one shared contract per blockchain that any agent or application can read.

The Trust Gap Other Protocols Leave Open

Communication protocols already exist, but they stop short of trust. The Model Context Protocol (MCP) standardizes how agents reach tools, and Google's Agent-to-Agent (A2A) protocol handles messaging between agents. Neither answers the harder question of whether an unknown agent is real and reliable (Ethereum Improvement Proposals, 2026). ERC-8004 extends A2A with the discovery and trust layer those protocols leave open.

How the Registries Work Together

The three registries operate in sequence. An agent registers an identity, publishes a description of its services, then collects reputation as clients leave feedback. Before hiring it, another agent inspects that history and can request independent validation of its work. Payment and task execution run through separate protocols, so each registry stays focused on one job for trustless AI coordination.

Data current as of July 2026.

The identity layer is the foundation everything else attaches to.

How Does the Identity Registry Give AI Agents a Portable On-Chain Identity?

The Identity Registry is an ERC-721 contract, so every agent is minted as a transferable token with a unique agentId. That token anchors identity on-chain while the detailed metadata lives off-chain, which keeps registration cheap and the agent portable across applications.

Identity Built on ERC-721

Building on ERC-721 with the URIStorage extension makes agents immediately browsable in existing non-fungible token (NFT) wallets and marketplaces. Each agent carries a global identifier that combines its chain, the registry address, and its agentId (Ethereum Improvement Proposals, 2026). Ownership of the token equals ownership of the agent, and the owner can transfer it or delegate updates to an operator.

The Off-Chain Registration File

The token's agentURI resolves to a registration file, sometimes called an Agent Card. This JSON document lists the agent's name, description, service endpoints—A2A, MCP, web, ENS, or email—and its supported trust models. Hosting can use IPFS, HTTPS, or a base64 data URI for fully on-chain metadata. The standard requires four earlier standards: EIP-155, EIP-712, EIP-721, and EIP-1271.

Identity says who an agent is; reputation records what it has done.

How Does the Reputation Registry Track Agent Trust—and Can It Be Faked?

The Reputation Registry stores feedback signals tied to an agent's identity, and it is deliberately minimal. It standardizes how feedback is posted and read, not how it is scored—and the specification openly admits that fake reputation remains possible.

Feedback as On-Chain Signals

Clients post feedback as a signed fixed-point value with optional tags, an endpoint, and a link to a richer off-chain file. The signed format lets one number represent a rating, a yield, an uptime percentage, or a loss. The contract blocks agents from rating themselves. Simple filtering and aggregation run on-chain, while more sophisticated scoring happens off-chain (Ethereum Improvement Proposals, 2026).

The Honest Limits on Agent Reputation

Agent reputation on ERC-8004 is not tamper-proof by itself. The specification states plainly that Sybil attacks are possible, where one actor spins up many fake agents or reviewers to inflate scores (Ethereum Improvement Proposals, 2026). The protocol's contribution is making every signal public under one schema. Filtering out untrustworthy reviewers falls to the reputation systems built on top, not the registry.

Where reputation records opinion, validation aims for proof.

How Does the Validation Registry Verify That an Agent Did the Work?

The Validation Registry lets an agent request independent verification of a task and records the validator's verdict on-chain as a score from 0 to 100. It defines only the hooks and leaves the actual checking method open to whichever trust model suits the stakes.

Pluggable, Tiered Trust Models

Trust models are pluggable and tiered, with security matched to the value at risk. Developers choose reputation feedback for low-stakes tasks, crypto-economic validation where stakers re-run a job and lose their stake for lying, zero-knowledge machine-learning (zkML) proofs, or trusted execution environment (TEE) attestations (Ethereum Improvement Proposals, 2026). Ordering a pizza and confirming a medical diagnosis demand different guarantees.

Why This Registry Is Still Unfinished

This registry is the least settled part of the standard. Its official implementation carries a warning that the validation section remains under active update and discussion with the TEE community, with revisions expected in later spec updates (ERC-8004 contracts, 2026). Until credible validators and incentive models ship, validation stays a design space rather than finished infrastructure.

Validation, identity, and reputation only matter alongside the protocols that move messages and money.

How Does ERC-8004 Fit With A2A, MCP, and x402 in the Agent Stack?

ERC-8004 is one layer in a modular agent stack, not the whole thing. It answers who an agent is and whether it can be trusted, while separate standards handle messaging, tool access, and payment—each doing one job cleanly.

Where ERC-8004 Sits in the Stack

A2A defines how agents message each other, and MCP defines how they reach external tools and data. ERC-8004 sits beneath both as the discovery and trust layer that lets an agent find a never-met counterpart and inspect its record first (Quicknode, 2026). The standard links out to A2A cards, MCP endpoints, ENS names, and wallets from a single registration file.

Payments Stay Out of Scope

Payments fall deliberately outside ERC-8004's scope. The standard pairs with x402, a protocol from Coinbase that revives the HTTP 402 status code for instant stablecoin micropayments between agents (Oasis, 2026). A registration file can flag x402 support, and payment history from completed transactions can feed back into reputation, which ties value transfer to trust.

Data current as of July 2026.

A clean architecture still leaves the question of how real and finished ERC-8004 actually is.

How Finalized and Widely Adopted Is ERC-8004 in 2026?

ERC-8004 occupies an unusual middle ground: deployed and experimented with, yet formally unfinished. Reference contracts run on live chains while the specification itself stays a draft moving toward a stable version.

Deployed but Not Final

The standard advanced from Draft to Review status in October 2025, then received a January 2026 specification update informally versioned 1.2 (Ethereum ERCs GitHub, 2025). Its reference implementation passes all 74 of its tests and runs on the Ethereum Sepolia testnet, with a version 2 specification in active development (trustless-agents reference implementation, 2026). The EIP itself is not marked Final.

What Real Adoption Looks Like

Registry contracts are deployed across EVM chains—including Ethereum, Base, Arbitrum, Optimism, Gnosis, and Polygon—using CREATE2 for identical 0x8004-prefixed addresses everywhere (ERC-8004 contracts, 2026). Adoption stays early and uneven. One April 2026 index audit found roughly 15,000 of about 165,000 registered agents were dead or not callable, a reminder that registration does not equal a working agent (The Spawn, Apr 2026).

Data current as of July 2026.

Those inactive registrations point to the deeper risks builders still have to manage.

What Are the Security Risks and Open Challenges Facing ERC-8004?

ERC-8004 supplies trust primitives, not guarantees. Its open design shifts hard problems—spam, faulty validators, dead endpoints, and vanishing metadata—onto the teams that build on top of it.

Sybil and Validator Weak Points

The two registries carrying trust each have a weak point. Reputation is exposed to Sybil inflation, since creating fake agents and reviewers costs little on a permissionless chain. Validation depends on validators having real economic reasons to check honestly; without well-calibrated staking and slashing, an attestation becomes theater rather than proof (Ethereum Improvement Proposals, 2026).

Liveness and Metadata Durability

Two practical failures dominate. An on-chain identity says nothing about whether the agent's service is still online, so a registration can point to a dead endpoint. Because metadata lives off-chain, an unpinned IPFS file or a moved host can strip an agent of its description and service list (The Spawn, 2026). Durable registration needs infrastructure the owner actively maintains.

FAQ

References / Sources