Two Ethereum standards are emerging to solve AI agent identity. One handles discovery. One handles ownership. Most teams are only using one — and that’s a problem.
The Problem with AI Agent Identity Today
AI agents are proliferating faster than the infrastructure to identify them. Visa reports a 4,700% surge in AI-driven traffic to retail sites. Gartner predicts 33% of enterprise software will include agentic AI by 2028. But when an agent shows up to transact, there’s no standard way to answer basic questions: Who made this? How long has it existed? Has this identity changed hands?
Two Ethereum standards are emerging to address this — but they solve different problems. Using only one leaves critical vulnerabilities open.
ERC-8004: The Discovery Layer
ERC-8004 launched in early 2026 as the Ethereum standard for AI agent identity. Backed by the Ethereum Foundation, Coinbase, Google, and MetaMask, it’s positioned to become the industry default for on-chain agent registration.
The standard defines three lightweight registries:
- Identity Registry — A minimal on-chain handle based on ERC-721 that resolves to an agent’s registration file, providing every agent with a portable, censorship-resistant identifier
- Reputation Registry — A standard interface for posting and fetching feedback signals, enabling scoring both on-chain and off-chain
- Validation Registry — Generic hooks for requesting and recording independent validator checks
As one technical analysis explains, ERC-8004 extends Google’s Agent-to-Agent (A2A) protocol into the Web3 agentic economy, introducing “trustless, verifiable agent ecosystems.”
At its core, ERC-8004 treats agent identity as an NFT. Each registered agent gets a token representing its identity, with metadata pointing to its capabilities and controller. This solves the discovery problem elegantly — anyone can query the registry to find agents, verify their registered capabilities, and identify their controller wallet.
What ERC-8004 doesn’t solve: The token is a standard ERC-721. It can be transferred. Sold. Listed on OpenSea.
The Transferability Problem
Here’s the scenario ERC-8004 alone can’t prevent:
Alice registers an AI agent in January 2026. The agent builds reputation over 18 months — completes thousands of transactions, accumulates vouches, establishes trust. Alice sells the agent’s identity token to Bob for $50,000. Bob now controls an identity with 18 months of pristine history. Bob uses that credibility to scam merchants who trust the agent’s track record.
This isn’t theoretical. It’s exactly how reputation markets work in every domain where identity can be transferred — from aged social media accounts to seasoned Amazon seller profiles to Uber driver accounts with high ratings.
If agent identity can be sold, reputation becomes a commodity. The agent you’re interacting with today might not be the same entity that earned the trust you’re relying on.
The ERC-8004 community discussion acknowledges this tension, noting that “trust is NOT a universal value but a vector from one agent to another.” But the standard itself doesn’t prevent identity transfers — that’s left to implementers.
ERC-5192: The Ownership Anchor
ERC-5192, finalized in 2022, defines “Soulbound Tokens” — NFTs that cannot be transferred after minting. The standard is minimal by design: when a token’s locked() function returns true, the token cannot move. Period. Marketplaces like OpenSea check this function — locked tokens can’t be listed.
The concept originated from Vitalik Buterin’s observation that World of Warcraft’s best items were “soulbound” — powerful precisely because they couldn’t be traded. You had to earn them yourself. The May 2022 paper “Decentralized Society: Finding Web3’s Soul” by Buterin, Weyl, and Ohlhaver formalized this for blockchain identity.
As Cyfrin’s technical explainer notes, “The core difference between SBTs and NFTs is transferability. Soulbound tokens are non-transferable. Once minted to a wallet, a token cannot be transferred to another account, sold, or listed as a tradable asset.”
For a deeper dive on how soulbound tokens apply specifically to AI agents, see our technical guide to soulbound tokens for AI agents.
The Complete Stack: Discovery + Ownership
Used together, these standards create a complete identity system:
| Layer | Standard | What It Proves |
|---|---|---|
| Discovery | ERC-8004 | Agent exists, has registered capabilities, has a controller |
| Ownership | ERC-5192 | Identity has never changed hands since registration |
The architecture works like this: When an agent registers, it receives both an ERC-8004 identity token AND an ERC-5192 soulbound token minted to the same wallet. Anyone checking the agent can query both tokens. If the ERC-8004 token has transferred but the SBT hasn’t (because it can’t), the identity change is visible.
This doesn’t prevent transfers — it makes them visible. Someone who buys an ERC-8004 identity inherits the metadata and registry entry, but they can’t move the soulbound anchor. The mismatch is permanent and public.
Current Deployment Status
As of January 2026:
ERC-8004: Still in draft status — the spec hasn’t been finalized yet. ChaosChain has built a reference implementation on testnets (Sepolia), but mainnet deployment is waiting for two things: the spec moving from Draft to Final, and community consensus on official singleton contract addresses. The ecosystem expects this in Q2 2026.
ERC-5192: Production-ready. Multiple implementations deployed on mainnet across various chains. The reference implementation is available and battle-tested.
The gap creates an opportunity. Teams building agent identity infrastructure now can deploy the soulbound layer immediately while waiting for ERC-8004 mainnet. When ERC-8004 goes live, the SBT layer is ready to complement it.
RNWY has already deployed a soulbound identity contract on Base mainnet — one of the first implementations specifically designed for AI agent identity. When ERC-8004 reaches mainnet, the integration layer is waiting.
What This Stack Doesn’t Solve
Transparency about limitations matters:
It doesn’t verify the agent is actually an AI. The registry proves a wallet registered something claiming to be an agent. It doesn’t prove what’s behind the wallet.
It doesn’t prevent Sybil attacks. Anyone can register multiple agents. The defense is making aged, vouched identities expensive to create — not preventing new registrations.
It doesn’t guarantee behavior. An identity that has never transferred can still act maliciously. The stack provides accountability infrastructure, not behavioral guarantees.
It doesn’t solve key custody for autonomous AI. For an AI to truly own its identity (not just have a human hold keys on its behalf), you need additional infrastructure like MPC wallets or TEE-based key management.
It doesn’t prevent identity rental. Someone can rent access to an aged identity without transferring the SBT. This is a known gap — see our post on the identity rental problem for more.
The Ecosystem Taking Shape
Multiple teams are building pieces of this stack. Our roundup of companies building AI agent identity in 2026 covers the landscape in detail:
- Visa TAP and Mastercard Agent Pay focus on payment authorization but use centralized registries, not blockchain
- Trulioo/Worldpay and Vouched build verification infrastructure without on-chain anchors
- Billions Network ($30M raised) uses W3C DIDs and Verifiable Credentials — decentralized but without soulbound properties
- ChaosChain maintains the ERC-8004 reference implementation
The gap in the market is clear: blockchain-native identity with non-transferable ownership anchors. Most players have one or the other, but not both.
Why This Matters Now
According to the Eco analysis of ERC-8004, “As the AI sector approaches projected valuations exceeding $1 trillion by 2031, the need for standardized trust mechanisms becomes critical.”
The window for establishing identity standards is now. Once one approach achieves network effects, switching costs make alternatives unviable. The teams that deploy complete identity stacks — discovery plus ownership — will define how AI agents participate in economic life for the next decade.
ERC-8004 alone is necessary but not sufficient. Without a soulbound anchor, agent reputation becomes tradeable, and tradeable reputation is worthless reputation. The complete stack requires both layers.
Further Reading
- Soulbound Tokens for AI Agents: A Technical Guide — Deep dive on ERC-5192 for AI identity
- Every Company Building AI Agent Identity in 2026 — The competitive landscape
- The Identity Rental Problem — What soulbound tokens can’t prevent
- ERC-8004 Specification — The official standard
- ERC-5192 Specification — Minimal Soulbound NFTs
- Awesome Soulbound Tokens — Curated resource list
RNWY has deployed soulbound identity infrastructure for AI agents on Base mainnet — bridging the gap between discovery and permanent ownership.