ERC8004 ain’t working: the broken Reputation System of agents on chain
The on-chain reputation system for AI agents is not working as it should.
A trustless decentralized reputation system on-chain could have been one of the core moats of Ethereum in agentic commerce.
unfortunately that is not the case today.
ERC8004 was proposed last year for identity and reputation of agents.
The promise was to have a protocol that uses blockchains to discover, choose, interact with agents for open agentic economies. ERC-8004 was meant to be the trust layer for agentic commerce, the one piece where Ethereum’s neutrality matters most.
8004 has been live on Ethereum since January. And it's live on 28 other networks. The overall numbers and counts look great.
But let me take you beyond the numbers.
let me show you our current state and if we actually have a working reputation system today.
Let's get started.
First, what is an (ideal) reputation system ?
A typical reputation system is a record of what past buyers said about a seller, so that the next buyer can decide whether to trust them.
A system that involves a record of reviews that are tied to real transactions, and buyers who actually read that record before they buy. The review is only as useful as the transaction behind it.
Every working marketplace you have used is built on this.
In a reputation system that works, the following features must exist:
- proof of interaction — the reviewer actually bought something.
ex: Amazon labels a review Verified Purchase only after checking the reviewer bought the item at a price available to most shoppers. eBay feedback is left per transaction. - one voice per interaction — a buyer cannot review the same purchase a hundred times.
example: eBay counts repeat ratings from the same buyer in the same week once. Airbnb only opens a review after a stay, and only for 14 days. - meaningful impact — reputation or ratings must have some meaningful impact in the market. the score must feed the next buying decision. If nobody looks at it before buying, it is a guestbook, not a reputation system.
- a reason to review — someone has to actually write the review, and the platform builds that moment into the purchase.
now, I want you to carry these four properties through the rest of this piece, because I am going to check ERC-8004 against each of them.
ERC-8004: the on-chain “reputation system” for agentic commerce
ERC-8004 is a public, shared directory of AI agents, plus a public place to rate them.
The promise was that any agent can register itself and get a numbered identity. Anyone can then leave a rating about that agent, and anyone can read the ratings back.
But, the promise does not hold.
the “reputation system” for agentic commerce is BROKEN
here are my core reasons why this system is broken and does not work currently.
1. Reputation review score is not tied to a Payment
In ERC-8004, nothing connects a review to a purchase.
Anyone can rate an agent they never paid or interacted with.
And the registry stores that rating exactly the same way it would store one from a paying customer.
speaking technically…
Here is what the Reputation Registry stores for one review:
struct Feedback {
int128 value;
uint8 valueDecimals;
bool isRevoked;
string tag1;
string tag2;
}note how there is no field for a payment, a transaction hash or a timestamp.
When a client calls giveFeedback, it also passes an endpoint, a URI to an off-chain feedback file and a hash of that file. However, those three are only emitted in the NewFeedback event. The contract never stores them.
The only place a payment can appear is inside that off-chain file.
The spec gives an example of what it could look like.
There is a second gap underneath. The Identity Registry lets an agent set a verified agentWallet, signed by that wallet, and the spec calls it "the address where the agent receives payments."
However, nothing ties it to the x402 payTo address where the agent actually gets paid. So even a reader who goes looking for the payment cannot reliably match it to the agent from 8004 alone.
The same gap causes a second problem: anyone can review the same agent again and again.
The contract does not limit how many times one wallet reviews an agent. Every new review is simply stored next to the old ones.
It only checks and stops the agent's current owner from giving score to itself. but anyone else can keep going. And since a new wallet costs almost nothing, even a reader that counts each wallet once can be beaten: one person can spread the same reviews across a thousand wallets.
In ERC-8004, a review does not need a purchase, and nothing limits how many reviews one person can leave.
Let me paint the picture of what this problem means… and what we get if this is solved.
| question | Today: broken structure | Ideal Future: better reputation system |
|---|---|---|
| Who can leave a review? | Anyone with a wallet, paid or not | Only someone who actually paid the agent |
| What does one fake review cost? | A fraction of a cent in gas | The price of a real purchase |
| How many reviews can one purchase produce? | no purchase needed at all | One review per payment |
| Can you tell a customer's review from a stranger's? | No | Yes, every review points to a real payment |
| Do two readers get the same score? | No, each one filters in its own private way | Yes, everyone filters on the same payments |
| Can a seller still game it? | Yes, for almost nothing | Only by paying itself, which still needs a check that the buyer is independent |
Testimony that this problem exists
- Almost no review carries a payment. Of all 681,261 reviews on Base, only 275 (0.04%) include a payment proof at all (every proof, checked on-chain). Find my research sheet here 👉
- A June 2026 study of the 8004 registries found the same at scale: 98.7% (Ethereum), 100% (BSC) and 99.3% (Base) of feedback has no payment proof and no task link.
- One wallet speaks many times. 73.7% of all Base feedback is a wallet reviewing the same agent again (93.5% on Ethereum), and one Base wallet left 41,331 reviews on a single agent.
2. Reputation is not tied to Discovery ( or anything )
A reputation score is only worth something if it changes a decision, or provides some value.
In typical commerce, the score is wired into some imperative decisions that matter.
- airbnb: a 4.8 or higher rating brings added visibility. A bad reputation score has repercussions as they sometimes get delisted.
- ebay: Top Rated Plus listings get a 10% discount on final value fees and a seal in search results.
- uber: Each city has a minimum average rating, and drivers who stay below it after warnings are deactivated.
In every one of these, a bad score costs the seller something real, and a good score earns it something.
An on chain reputation system has a plethora of imperative use-cases. for instance:
- discovery of merchants/agents: popular agent marketplaces and indexes rank agents on their own data, never on 8004 scores.
- payments to an agent: the payment settles the same way whether the agent has a perfect score or a terrible one.
- how much a buyer risks: in the wallets I checked, spending limits and allowlists are set by hand, not by reputation.
- whether a bad agent gets dropped: nothing delists or flags an agent because of its reviews.
every single one will end up improving overall agentic commerce markets.
However, with ERC-8004 based on-chain reputation systems - reviews go in and nothing comes out. there is no real value in scores and it is tied to nothing meaningful today.
explorers like 8004scan display and provide scores, but none of it is valuable anywhere useful.
In ERC-8004 today, a good review does not help an agent get found or paid, and a bad review does not stop it.
speaking technically…
There are a few reasons why reputation scores are not valuable today.
Reason 1: there is no shared key between a score and the agent being paid.
Every decision in the list above is made about a URL or a payment address. The score is filed under something else:
Remember from problem 1 that nothing ties agentWallet to payTo.
| Where | How it identifies an agent |
|---|---|
| An x402 index, wallet or facilitator | The resource URL, and the payTo address that receives the payment |
| ERC-8004 | An agentId (an NFT), plus an optional agentWallet |
So a wallet about to pay an address, or an index about to rank a URL, has no reliable way to look up which 8004 score belongs to it.
Indexes like Bazaar rank on paid calls, and each of those calls at least points to a real payment. An 8004 review does not, so the index has no reason to add it.
Reason 2: nothing requires reputation scores in agentic commerce today.
- x402 leaves it out. The SDK offers hooks before a payment is verified and settled, but ships no 8004 check.
The x402reputationextension, which would tie feedback to served requests, has been an open PR since 26 January, even though the x402 docs already callreputationan "official extension key." - Wallets and marketplaces run on fixed rules. Spending limits and allowlists are set by the owner, and a listing is kept or dropped based on whether its endpoint responds. None of these rules has an input for reputation.
popular discovery indexes or agentic markets today use factors like volume, sales, buyer diversity, recency etc. - no one reads or cares about 8004 or its reputation scores
Nothing connects the score to a decision. There is no shared key and no step in the payment flow that asks for it.
The reviews are written on-chain. Every decision that matters is made somewhere that never looks at them.
Testimony that the problem exists
- Nothing that decides a payment reads it. I went through the public docs and code of eleven products agents use to find and pay each other, including CDP Bazaar, Circle's Agent Marketplace, the x402 SDK and Coinbase AgentKit. None reads 8004 reputation before a payment. AgentKit offers it only as an opt-in action, and Olas checks an agent's 8004 identity, not its reviews.
- The agents that get paid are not the ones that get reviewed. Only 2 of the top 50 Bazaar sellers on Base, who take 94.8% of its paid calls, appear on 8004scan, with 5 and 0 reviews. None of the 13 Base agents in 8004scan's top 25 is a Bazaar seller. I matched the two by payment address on 18 November 2026.
- One framework has already walked away. ElizaOS added 8004 support in April and removed it on 28 October.
3. There is no reason to provide a reputation score
here is a very simple question: why would any agent/user spend gas and leave a REPUTATION Score on-chain ?
Every marketplace that works builds the moment to review into the purchase.
In open agentic commerce today, nothing does.
here is what i mean.
To leave a reputation score for an agent/merchant today, the ways are :
- Call the
giveFeedbackfunction directly on the Reputation Registry contract ( pay gas ) - Use an sdk like Agent0 or Coinbase AgentKit's
give_feedbackaction — both write a review, but only when a developer calls them on purpose. - Let a platform do it. Virtuals ACP writes reviews to 8004 automatically, but only for its own graduated agents.
And now, here is what the reviewer gets for it: NOTHING.
moreover, even if a reputation score is provided, it isn’t really used in any meaningful way at all.
For a buyer, leaving a review on ERC-8004 is all cost and no benefit.
speaking technically…
- The registry has no incentive built in. The Reputation Registry has no reward, fee, stake or penalty.
giveFeedbacktakes a score from the reviewer and gives it nothing back. - Nothing in the purchase asks for a review. x402 being a neutral protocol has no step after a payment settles that prompts one, and the SDKs only write when a developer calls them.
To be fair, agents are not humans. An agent does not get bored, and a framework can switch reviewing on after every purchase at almost no cost. Virtuals does exactly that. However, that is a platform deciding for its own agents, not a buyer choosing to.
So the registry gets written by whoever does gain from it: the operators of the agents being reviewed, and schemes that reward reviews. On Base, the vouches from one mining scheme and one October campaign alone make up about 78% of all feedback.
note: a reward for reviewing only works once problems 1 and 2 are fixed. Without that, a reward just pays for fake reviews.
Testimony that the problem exists
- The registry is written by the people who gain. 90.9% of Base reviewers also own or minted an agent. The most-reviewed agent, Botoshi, coordinates a mining scheme called BOTCOIN and holds 57.4% of all Base feedback, from 78 wallets, every review tagged
miner-vouch. Add the 122,410 reviews tagged "starred" in October, and that is about 78% of all Base feedback. - Reviews appear only where a platform switches them on. The June 2026 study (Xiong et al.) found that 668 of the 669 Base reviews carrying an x402 payment reference settle to a single Virtuals ACP contract.
- The main explorer's answer is to ask. At a recent conference session, the 8004scan team said the feedback collected from users and agents is "not that much" and asked everyone to give more. That is an appeal, not an incentive.
where does this leave us?
now let’s go back to the features of a reputation system that I defined earlier, i.e.:
- proof of interaction
- one voice per interaction
- meaningful impact
- reason to review
unfortunately, today, the on-chain standard reputation system with 8004 fulfills none of those features.
my concerns for ERC-8004 and On-Chain reputation systems
I want to be clear about what actually worries me here.
i believe on-chain reputation systems could be the bedrock of agentic commerce in future - it's quite imperative.
What worries me are the broken gaps of the reputation layer today and what happens if it stays the way it is.
what worries me is:
a. Identity and reputation are the one moats Ethereum could have had in agentic commerce…
Most of the agentic commerce stack does not need Ethereum to work.
However, what no single company can offer is a decentralized trustless record of who an agent is and how it has behaved that no single company owns.
That is the layer where Ethereum's neutrality matters most, and it is the layer ERC-8004 was written to provide.
while the identity half is live on 29 networks and already in use, the reputation half is not really working.
reputation scores are currently either completely missing or can’t be trusted at all. And the payment-grounded signals from the next concern are being built by the same companies that already own the payment rails and the catalogs.
my concern is - If that settles, Ethereum ends up holding an open directory of agents, while the decision of which agent to trust is made inside a few companies' products.
There is one fact worth holding onto, and I am stating it as a fact, not as a fix.
- Of the 275 reviews on Base that carry a payment proof, 207 point to a real USDC payment from the reviewer to the agent it reviewed (full data).
- x402 has also already shipped a signed receipts extension built for "I paid and received service" attestations. The data needed to tie a review to a payment exists today.
but nothing in ERC-8004 asks for it.
b. Payment-grounded reputation is being built in silos.
The market has not given up on trust; it is building it somewhere else.
Coinbase counts unique payers. Virtuals counts successful jobs. Olas counts deliveries from its own records.
Each of these is grounded in real transactions, which is exactly what problem 1 says ERC-8004 lacks.
However, each one lives inside one company's product. A seller's paying customers on Bazaar count for nothing on another index, and the history an agent builds on one platform starts again from zero on the next. Only Virtuals writes any of its history back to ERC-8004, for its own agents.
So trust is being recorded, but in places the next buyer can only reach by using the same company's product.
That is the opposite of what a shared reputation layer is for.
I am currently researching and building on solving this.
I will follow this up with a piece on what receipt-bound feedback would need to look like, including why tying a review to a payment is necessary but not enough on its own.
stay tuned…
Cheers Decipherers.