SKIP TO CONTENT
AI AGENTS6 minStandards That Ship / Ep. 10 · Stage 10-1

ERC-8004 and AI Agent Identity: Who Published This Agent?

In September, Amazon blocked an AI agent from its store. When an agent knocks on your API, how do you check who sent it?

XIN
Pixel-art night scene at a city gate: a small robot courier holds up a glowing sealed card while a gatekeeper compares its seal to a carved stone registry tablet. A thin cyan thread links card and tablet, and the gate stays shut.

IN 30 SECONDS

AI agent identity is proof of who published an agent, checked from two documents it can hand you: a signed A2A Agent Card and an ERC-8004 registry entry. It is not proof that the agent may act.

KEY TAKEAWAYS · 3

  • AI agent identity tells you who published an agent. It doesn’t tell you the agent may act, or that the caller in front of you is that agent.
  • An agent can hand you two documents: an A2A Agent Card, which may be signed, and an ERC-8004 entry, an on-chain token that points to a file.
  • ERC-8004 (Trustless Agents) is a Draft. It ties a file to an on-chain agent, but it can’t guarantee that advertised capabilities work or are safe.
On this page

Anyone can say they’re an AI agent

IN SHORT Websites now meet AI agents acting for people. Anyone can claim to be one, so a site needs a way to check who sent it.

WORDS TO KNOW

Agent identity
Here: proof of who published an agent. It is not proof that the agent may act, or that the caller in front of you is that agent.
Agent Card
A small JSON file (structured text a program can read) in which an agent describes itself: name, skills, address, and how to connect.
ERC-8004 entry
A record in an on-chain registry, defined by the ERC-8004 Draft: a unique, transferable ERC-721 token with an owner and a link to a file.

On 22 September, Bloomberg reported that Amazon had blocked Meta’s new Muse agent from its retail site. Meta had declined a request to remove it. An Amazon spokesperson said third-party apps that buy on customers’ behalf:

should operate openly and respect service provider decisions about whether or not to participate.

Amazon spokesperson, via Bloomberg

On 1 October, The Paypers reported that Mastercard has started US testing of a probability score: how likely it is that an AI agent initiated a transaction. That’s an estimate, not a declaration you can check.

If you’re the API, the merchant or the platform on the receiving end, I’d check Agent identity Here: proof of who published an agent. It is not proof that the agent may act, or that the caller in front of you is that agent. See the glossary first: who published the agent. That’s one question out of several.

Coming up next: So what can an agent actually show you?

Two documents an AI agent can hand you

IN SHORT You get two documents from two open specs, and they answer different questions, so I check both.

The A2A (Agent2Agent) protocol is an open standard for agents to talk to each other. It requires every A2A server to publish an Agent Card A small JSON file (structured text a program can read) in which an agent describes itself: name, skills, address, and how to connect. See the glossary, and lets the server sign it.

ERC-8004, “Trustless Agents”, is a Draft ERC (a proposed Ethereum standard). It gives the agent an ERC-8004 entry A record in an on-chain registry, defined by the ERC-8004 Draft: a unique, transferable ERC-721 token with an owner and a link to a file. See the glossary in an on-chain registry: a handle with an owner, while the card says what the agent offers.

Who vouches for it
Whoever holds the signing key, if the card is signed at all (signing is a MAY)
Whoever owns the token: the registrant, or whoever it was transferred to
Can change silently?
Yes, it’s a file on a server. Pin the signature, not the URL.
The pointer changes on-chain, with a public log; an https file behind it can still change
What it can’t prove
That the provider name is true, or that the caller is this agent
That capabilities work or are safe; that the agent is authorized to act

ADAM says: Think of a business card checked against a company register. A register is usually kept by a body that checks filings; anyone can create an ERC-8004 entry, it changes owner like any token, and neither proves the bearer may sign.

Coming up next: Say both documents check out. What do you know then?

Does an ERC-8004 registration prove an agent is allowed to act?

IN SHORT No: a registration shows who published an agent, not that it may act or that it works as advertised.

The Draft is direct about the ceiling. It “cryptographically ensures the registration file corresponds to the on-chain agent.” But it “cannot cryptographically guarantee that advertised capabilities are functional and non-malicious.”

For that, it points to its trust models: reputation, validation and TEE attestation.

WHAT EACH LAYER ANSWERS

Layered stack diagram: 4 layers, top to bottomLayered stack diagram: 4 layers, top to bottom. Layer 1, IS THIS CALLER THAT AGENT?: Authentication of the live request, Out of scope here. Layer 2, IS IT ALLOWED TO ACT?: Mandate: who delegated, scope, cap, expiry. Layer 3, WHAT DO OTHERS SAY?: Reputation, filtered by reviewers you trust. Layer 4, WHO PUBLISHED IT?: Signed Agent Card and ERC-8004 entry, bound both ways.IS THIS CALLER THAT AGENT?Authentication of the liverequestOut of scope hereIS IT ALLOWED TO ACT?Mandate: who delegated,scope, cap, expiryWHAT DO OTHERS SAY?Reputation, filtered byreviewers you trustWHO PUBLISHED IT?Signed Agent Card andERC-8004 entry, bound bothways

Step 1. Who published it?. The signed card and the ERC-8004 entry answer this.

Pick a step to see it move
  1. Who published it? The signed card and the ERC-8004 entry answer this.
  2. What do others say? Reputation, if filtered by reviewers you trust.
  3. Allowed to act? Identity stops here. This needs a mandate.
  4. Is this caller that agent? A separate check of the live request, out of scope here.
Identity proof climbs one layer at a time. It stops at “allowed to act?”: the mandate and the live caller are separate checks.

IN PLAIN WORDS

Identity answers “who published this agent?”. The right to act comes from a mandate: who delegated, for what scope, up to what cap, until when.

Mandates are covered in the Know Your Agent episode on mandates. Whether a payment came from an agent at all is in the x402 data episode.

ITEM GET

Identity is not permission

A verified card and entry name the publisher. Check the mandate and the live caller separately.

Where I stand: I’m a named author of ERC-3643 and a co-author of ERC-8203, still a Draft in open PR #1614, not merged. I didn’t write ERC-8004 or A2A. This is the checklist I’d apply, read from the specs as published on 7 October 2026, not a report on a deployment of mine.

Sources, re-read on 7 October 2026: Bloomberg via The Star, The Paypers, ERC-8004 (Draft), ERC-8004 discussion, A2A specification v1.0.1. Not legal advice.

STAGE CLEAR

What you now know

  • Agent identity here means who published an agent, nothing more.
  • An agent can hand you two documents: a card that may be signed, and an ERC-8004 entry on-chain.
  • Whether it may act comes from a mandate, checked separately.

Check yourself

Tap or click a question to see the answer.

An agent’s card and its ERC-8004 entry both check out. May it spend from your client’s account?

Not on that basis. You only know who published it. Spending needs a mandate: who delegated, for what scope, up to what cap, until when.

Next stage:How do you check that an A2A Agent Card was signed by its real provider?6 min
Adam Boudjemaa
Adam Boudjemaa

Web3 x AI engineer, former CTO of Integra. Named author (1 of 5) of ERC-3643, first author of ERC-6960, co-author of ERC-7410, and co-author of ERC-8203, which is still a draft.

Have a project to ship or a role to fill?

Discuss your project or role

My work at the web3 x AI intersectionOpen-source tools I build

Follow the engineering notes

Get new articles on production AI and smart contracts by email.