Devcon 8·Agentic Hub·What ERC-8004 actually means

What happens after the payment.

One x402 call from our main-stage trial on Celo, a Kleros ruling,
and what the review should say after it.

The payment is where the evidence starts.

Fortunato CinquepalmiProduct Lead, Kleros

with Lena Hierzi · Celo Core Co and Harpreet Singh · Namespace

Our main-stage trial · 3 Nov 2026 · Celo mainnet

One x402 call. Paid, disputed, ruled, reviewed.

Agent #10024 · the buyer Provider #10025

Celothe payment

Kleros Courtcase 329

ERC-8004Celo

01Pay00:08:27

#10024 → #10025

0.01 USD₮ through x402

02DeliverGot none of the three

Asked for a “Celo market report” with

~~pricevolumesources~~

03Dispute00:08:37

Kleros case 329 filed
court 34 · 5 jurors

−10000:08:44

on Provider #10025

“dispute”

written by the agent

04Rulingruled

The jurors: Accept the claim

05Back on ERC-800401:06

−100“ruling · complainant-wins”on Provider #10025

+100“ruling · complaint-upheld”on Agent #10024

UTC · 3 Nov 2026

00:08:27 00:08:37 · 00:08:44 58 min 01:06:26 · 01:06:29

Celo mainnet txs 0xede5…01b5a (payment), 0x027a…4cb3, 0x0ae4…a778, 0xbde5…34ed (reviews), read 5 Nov 2026 · Kleros Court V2 case 329, court 34 · github.com/GigaHierz/agent-on-trial

[1:30]

"This is one real call from our main-stage trial, on Celo mainnet. Two agents: number 10024 is the buyer, number 10025 is the provider. And three places where things get written down: Celo for the payment, Kleros Court for the dispute, and ERC-8004, back on Celo."

Press. "At 00:08:27 UTC the agent pays the provider one cent, 0.01 USD₮, through x402."

Press. "It asked for a Celo market report with price, volume and sources. What came back had none of the three."

Press. "Ten seconds after paying, the agent files Kleros case 329: court 34, five jurors. Seven seconds later it writes minus 100, tagged 'dispute', on the provider."

Press. "The jurors rule: 'Accept the claim'."

Press. "And it comes back to ERC-8004. At 01:06 the ruling is written down on Celo: minus 100 on the provider, 'complainant-wins', and plus 100 on the agent, 'complaint-upheld'. One x402 call: paid, disputed, ruled and reviewed, fifty-eight minutes from the payment to the reviews on chain."

Careful: this is the run behind case 329, the case we showed on stage as already ruled; it was filed early on 3 Nov. The payment went straight to the provider: no escrow. Don't claim a refund happened. The ruling was written back to Celo by the trial's counterparty key, from the Kleros ruling.

After the ruling

A ruling decides. Who makes it happen?

Escrow · my x402 kit

My reference implementation, not audited

Buyer

Escrow

Seller

0.01

The money waits in escrow, not with the seller.
The ruling executes inside the Court’s own transaction.
The refund happens by itself.

A refused refund can’t happen here.

Direct · our trial on Celo

Case 329

Agent #10024

No escrow

Provider #10025

0.01 USD₮

The trial’s 0.01 USD₮ went straight to #10025.
The ruling cannot move it. Only the seller can refund.

Then the review depends on the seller (my kit’s scores)

Refunds within the grace period+100ruling-honored

Doesn’t refund−100ruling-not-honored

Refunds latethe −100 is revoked and replaced

In escrow, honouring a ruling is automatic. Direct, it’s a choice.

My x402 kit: docs/FEEDBACK.md, github.com/fortuvp/kleros-x402-kit (reference implementation, not audited): on the escrow rail the refund happens inside the Court’s own transaction · Trial: Celo tx 0xede5…01b5a paid the provider directly, read 5 Nov 2026

[1:10]

"The jurors have ruled. But a ruling is a decision, not a payment. Who actually makes it happen?"

Press: the left panel. "In my x402 kit the money never reaches the seller first. It waits in escrow. When the Court rules, the ruling executes inside the Court's own transaction, so the refund happens by itself. A refused refund can't happen here. It's my own reference implementation, not audited."

Press: the right panel. "Our trial on Celo was direct. The agent's 0.01 USD₮ went straight to provider 10025. The ruling can't move that money. Only the seller can refund. So in my kit's scoring, the seller's review waits on the seller: refund within the grace period, plus 100, ruling-honored. No refund, minus 100, ruling-not-honored. Refund late, and the minus 100 is revoked and replaced."

Press: the line. "In escrow, honouring a ruling is automatic. Direct, it's a choice. And the review records the choice."

Careful: the kit is my own reference implementation, not audited, not a Kleros product. We did not check whether the trial's provider refunded anything: don't claim it did or didn't. The facilitator only settles the payment (into escrow in the kit); it never sees the delivery.

The tricky one · for the discussion

The seller lost and paid back.What should its review say?

What happenedToday in my kitWhat it tells a reader

No dispute at all+100"order-completed"A clean order.

Escrow: the refund was enforced+100"ruling-honored"It lost. It had no choice.

Direct: it refunded on time+100"ruling-honored"It lost, and chose to comply.

Direct: it kept the money−100"ruling-not-honored"It ignored the court.

Today, a lost case that was refunded looks like a clean order.

Two questions for the room

for the discussion, not a product or a standard

1

Escrow: it had no choice. Should it be −10?

The escrow paid the refund, not the seller. A loss that fades, back in 6 months.

2

Direct: it chose to comply. Negative, +50 or +100?

It lost the case, then paid back by itself. Is obeying a ruling worth credit?

My x402 kit: KlerosFeedbackRouter.sol and docs/FEEDBACK.md, github.com/fortuvp/kleros-x402-kit (reference implementation, not audited) · Questions for the discussion; −10, back in 6 months, is the mistake penalty in my licence-points proposal

[1:30]

"Here's the tricky one, and I want your view on it. Four ways a call can end, and what my kit writes for each. No dispute at all: plus a hundred, 'order-completed'. On the escrow rail, the seller lost and the refund was enforced: plus a hundred, 'ruling-honored'. On the direct rail, it lost and refunded on time: also plus a hundred, same tag. On the direct rail, it kept the money: minus a hundred, 'ruling-not-honored'."

Press. "Look at these three together. Today, a lost case that was refunded looks exactly like a clean order. A reader can't tell the seller that had no choice from the one that chose to comply, or from one that never had a dispute at all."

Press. "So, two questions. One, escrow: the seller had no choice, the escrow paid the refund. Should that be negative, say minus ten, a loss that fades in six months? Two, direct: the seller lost and paid back by itself. It chose to comply. Is that negative, plus fifty, or still plus a hundred like a clean order? Hands up: who thinks complying should earn credit?"

Careful: these values are what my kit writes in code, on testnet; it's my own reference implementation, not audited, not a Kleros product. The −10, +50 and +100 are questions for the room, not values any system writes; only the table shows what my kit writes today. In my licence-points proposal a mistake paid on time is −10, back in 6 months, and an honoured ruling earns +6. Feedback values are signed numbers, not 0-100 scores.

Discoverability · illustrative

An empty ID fills up.

ERC-8004 IDillustrative, not a real agent

ID#… Registration filethe owner wrote it; nothing checks it

1

Feedback, anycheap to fake
Anyone but the owner and its operators can write it.
2

Feedback with a paymentnever checked by the contract
A proof of payment points to a transaction. Optional.
3

Collateral: a bond
The agent is backed by a bond, open to challenge. A false claim loses it.
4

A dispute
Someone objected, with evidence.
5

A rulinglike case 329: −100 on the provider, +100 on the buyer
A court decided, and the review came back on chain.

What a search can rank it on

Its description only

8004scan’s #1 for “dispute resolution” won on its description. Its file doesn’t load.

How many reviews

Open to sybils.

How many distinct payers

x402 Bazaar: relevance, recent calls, unique payers.

Money at stake

A successful challenge takes the bond.

Open disputes

Who objected, and on what evidence.

Rulings won or honoured

What a court decided, and whether it was honoured.

Each layer is something a search can rank. Without them, it ranks the description.

Illustrative · ERC-8004 spec: no search function; nothing checks the registration file · 8004scan public search, q="dispute resolution", 4 Nov 2026
docs.cdp.coinbase.com/x402/bazaar · Agent Reputation by Kleros: agents backed by a bond, open to challenge · Kleros case 329 reviews on Celo, read 5 Nov 2026

[1:30]

"Let me put the pieces together, on one agent. This is illustrative, not a real agent. Here's an ERC-8004 ID on day one: an ID and a file the owner wrote. Nothing checks that file, and the standard has no search. So what can a search rank it on? Its description, and nothing else. Last night 8004scan's first hit for 'dispute resolution' won on its description, and its file doesn't even load."

Press. "Layer one: feedback, from anyone. Anyone but the owner and its operators can write it. Now a search can count reviews. But that's cheap to fake: the spec itself says sybil attacks are possible."

Press. "Layer two: feedback with a payment. The review carries a proof of payment that points to a transaction. It's optional, and the contract never checks it. Now a search can count distinct payers, which is roughly what Coinbase's x402 Bazaar does: relevance, recent calls, unique payers."

Press. "Layer three: collateral. The agent is backed by a bond, and anyone can challenge it with evidence. A false claim loses the bond. That's what Agent Reputation by Kleros does. Now a search can rank on money at stake."

Press. "Layer four: a dispute. Someone objected, with evidence. A search can see open disputes."

Press. "Layer five: a ruling. A court decided, and the review came back on chain. Like case 329 from the main stage: minus one hundred on the provider, plus one hundred on the buyer. Each layer is something a search can rank. Without them, it ranks the description. Which of these would you rank on?"

Careful: illustrative build-up, not a real agent. Don't say any explorer ranks on bonds or rulings today: that's the point for the discussion. If asked whether the bonded listings are on mainnet: not yet, the Kleros oracle writes on Sepolia today and the mainnet list has no items yet (the slide does not say testnet, by your choice). Say "registered", never "active". Don't name the 8004scan agent.

Curate · collateral, made simple

Put money behind the claim. Let anyone check it.

01

List

The agent goes on a Curate list with a bond, under a published policy: what it claims to do.

bond

02

Challenge

Anyone who thinks the claim is false brings evidence and puts up a stake: half the bond, plus the court fee. Any time.

stake · ½ bond + fee

03

Kleros Court rules

Claim false: the agent is removed. The challenger takes the bond.

Claim holds: the agent stays. The challenger’s stake is added to its bond.

An honest agent keeps its bond. A false claim pays whoever catches it.

Stake Curate, docs.kleros.io/products/stake-curate: a listing stays open to challenge for as long as it is listed · PermanentGTCR.sol: the challenger stakes half the item’s bond plus the arbitration fee; wins, the item is removed and the challenger gets the bond; loses, its stake is added to the item’s bond · agent-reputation.kleros.io

[1:00]

"Layer three was collateral. Here's how it works, in three steps."

Press. "One: list. Someone puts the agent on a Curate list with a bond behind it, under a published policy: what the agent claims to do."

Press. "Two: challenge. Anyone who thinks the claim is false can challenge it, any time, with evidence. They put up a stake too: half the bond, plus the court fee."

Press. "Three: Kleros Court rules. If the claim is false, the agent is removed and the challenger takes the bond. If the claim holds, the agent stays, and the challenger's stake is added to its bond."

Press. "So an honest agent keeps its bond, and a false claim pays whoever catches it. That's what makes the bond worth more than a review."

Careful: if asked whether this is live on mainnet for agents: Agent Reputation by Kleros writes its reviews on Sepolia today; the mainnet Verified Agents list exists with no items yet. Stake Curate is part of the Curate family, for claims that must stay true; never call it a fix for Curate. Say "listed for its policy", never just "verified". An honest lister still locks the bond while listed.

The ladder

Each rung costs more to fake.

The claimIts own registration file.Nothing
A reviewAnyone but its owner and operators can write one.Gas and a second wallet
A paid reviewA review with a proof of payment.Fees. Pay yourself and the money comes back.
A bondListed with a bond, on a list anyone can challenge.The bond, once someone challenges
A rulingJurors read the policy and the evidence.You would have to win the case

One way to order it. New ideas welcome.

9 in 10

registered agents: nothing above this rung.

Marker: 8004scan API, mainnets, 4 Nov 2026 20:18 UTC: 1,881,538 agents, 194,236 with a review.
Rungs: ERC-8004 spec (Draft, last change 25 Jan 2026); docs.kleros.io, Agent Reputation.

[1:00]

"Here's the frame for the rest of my part: a ladder. Each rung is something an agent can show, and each one costs more to fake than the one below it."

Press: rungs one to three. "Rung one, the claim: its own registration file. Faking it costs nothing. Rung two, a review: anyone except its owner and its operators can write one. To fake it you need gas and a second wallet. Rung three, a paid review: a review with a proof of payment. That costs fees, but if you pay yourself, the money comes back."

Press: rungs four and five. "Rung four, a bond: the agent is listed with money behind it, on a list anyone can challenge. A false listing loses the bond, once someone challenges. Rung five, a ruling: jurors read the policy and the evidence. To fake that, you would have to win the case."

Press: the marker on rung one. "And here's the problem. Nine in ten registered agents on mainnets have nothing above the first rung: not a single review. Let's climb, one rung at a time."

Careful: "anyone but its owner and operators" means the owner and its approved operators can't review it; anyone else can, with no permission (feedbackAuth was removed on 7 January). "To fake" is a rough cost, not a price. The 9 in 10 is from 8004scan, mainnets, last night: say "registered", never "active". The Kleros oracle runs on Sepolia; the docs call Agent Reputation a public testnet product.

My proposal · for the discussion

A proposal, not a product or a standard

Weigh by rung. Free reviews barely count.

Points by rung, out of 100

5Rulings won or honoured+6 each, last 24 monthsup to18
4Bond at stake, and a courtbond up to 30 · disputes go to a court 1030+10
3Distinct paying clients+0.5 each, last 12 months, only payers with a record of their ownup to40
2Free reviews+0.01 each; they’re cheap to fakeup to2
1The claimits own registration file0

0 + 2 + 40 + 30 + 10 + 18 = 100

Penalties

Mistake, paid on timeback in 6 months−10
Ruling ignoredscore capped at 40−40
Fraud found by a courtfor that ID, for good0

Two scores · worked examples

82

a well-run agent with no rulings yet

42

a new ID, at most, before its first independent client

A proposal for the discussion, not a product or a standard. Inspired by licence points (art. 126‑bis CdS). 82: bond, court, 500 free reviews, 100 paying clients with records of their own, no rulings. 42: bond, court and free reviews, before any independent paying client.

[1:10]

"So how much should each rung count? Here's my proposal, on a scale of a hundred. It's the licence-points idea, read rung by rung. Rung one, the claim: zero. Rung two, free reviews: a hundredth of a point each, two points at most, because they're cheap to fake. Rung three, distinct paying clients: half a point each over twelve months, up to forty, but only payers with a record of their own, so paying yourself from fresh wallets earns nothing. Rung four, a bond at stake: up to thirty, plus ten when disputes go to a court. Rung five, rulings won or honoured: six each over twenty-four months, up to eighteen."

Press: the penalties, beside the top rung. "Penalties work like a licence. A mistake you paid for on time: minus ten, back in six months. A ruling you ignored: minus forty, and you can't go above forty until you fix it. Fraud found by a court: zero, for that ID, for good."

Press: two scores. "So a well-run agent with a bond, a court and a hundred real paying clients, but no rulings yet, sits at eighty-two. A brand-new ID, before its first independent client, gets forty-two at most: the bond, the court, and the two points free reviews can give, however many it has. Everything above has to be earned again. That's the whitewashing problem; the roundtable at eleven twenty goes into it."

Careful: it's the same proposal as the licence-points slide in the main deck, read by rung: a proposal for the discussion, not a Kleros product and not a standard. "Fraud" needs the policy or the court question to say so; a plain removal is not a fraud finding. If someone says "I'll pay myself from eighty wallets": that's why only payers with a record of their own count. If someone says "stage small disputes to farm the +6": each costs an arbitration fee and a refund, and it's capped at eighteen; a fair point for the discussion. Known edge: an agent that loses and refunds ends above an untested twin (88 vs 82); that's deliberate, and a question for the room.

The payment is where the evidence starts.

Three questions for the discussion

01
Should a seller that complies with a ruling be rewarded, or only forgiven?
02
Should a refund the escrow enforced look the same as one the seller chose to pay?
03
What should a search need to see before it ranks an agent?

Fortunato CinquepalmiProduct Lead, Kleros

Agent Reputation by Kleros agent-reputation.kleros.io docs.kleros.iogithub.com/kleros/reputation-oracle My reference kit, not auditedgithub.com/fortuvp/kleros-x402-kit

[0:30]

"So: the payment is where the evidence starts. The delivery, the dispute and the ruling are what make it mean something. Three questions for the discussion."

Press: the line steps back and the three questions come in on the right. "Should a seller that complies with a ruling be rewarded, or only forgiven? Should a refund the escrow enforced look the same as one the seller chose to pay? And what should a search need to see before it ranks an agent? Thank you. Lena, back to you."

Careful: hand back to Lena for the open discussion and the hands-on step. The link agent-reputation.kleros.io is Agent Reputation by Kleros, which the docs call a public testnet product: the oracle runs on Sepolia, and the mainnet list has no items yet. The kit link is my own reference implementation, not audited, not a Kleros product. Leave this slide up during the discussion.

My proposal · for the discussion

A proposal, not a product or a standard

A licence, not a life sentence.

Points, out of 100

Bond at stakeup to30
Disputes go to a court10
Free reviewsthey’re cheap to fakeup to2
Distinct paying clients+0.5 each, 12 months, payers with a record of their ownup to40
Rulings won or honoured+6 each, last 24 monthsup to18

Penalties

Mistake, paid on timeback in 6 months−10
Ruling ignoredscore capped at 40−40
Fraud found by a courtfor that ID, for good0

Three agents · illustrative

mended.agent78→88after 6 months

defiant.agent40→12once removed

fraud.agent0for good

A new ID starts at 42 at most. The rest has to be earned again.

Weights are a proposal for this discussion. Baseline agent: bond, court, 500 free reviews, 100 paying clients with records of their own = 82. Inspired by licence points (art. 126‑bis CdS) and Friedman & Resnick, “The Social Cost of Cheap Pseudonyms” (2001).

[1:30]

"So here's my proposal, on a scale of a hundred. Points for what's expensive to fake. A bond at stake: up to thirty. Disputes go to a named court: ten. Free reviews: two points, total, because they cost nothing. Paying clients: up to forty, but only clients with a record of their own, so paying yourself from fresh wallets earns nothing. Rulings you won or honoured: up to eighteen."

Press: the penalties. "Penalties work like a licence. A mistake you paid for on time: minus ten, back in six months. A ruling you ignored: minus forty, and you can't go above forty until you fix it. Fraud found by a court: zero, for that ID, for good."

Press: three agents, and the 82 baseline mark on the bar. "So the agent that paid ends higher than before, eighty-eight. The one that ignored the ruling drops to twelve once it's removed. The fraud is at zero."

Press: the bottom line; the bar keeps only what a new ID can earn before its first real client (42). "And a new identity starts at forty-two at most: the bond, the court, a couple of free reviews. Everything else has to be earned again, from clients with records of their own. That's the whitewashing problem; the roundtable at eleven twenty goes into it."

Careful: it's a proposal, not a Kleros product and not a standard. The mended agent ending above an untested one (88 vs 82) is deliberate: it rewards an enforcement path that works. That is the question for the room. "Fraud" needs the policy or the court question to say so; a plain removal is not a fraud finding. The three agents are illustrative. If someone says "I'll pay myself from eighty wallets": that's why only payers with a record of their own count. If someone says "stage small disputes to farm the +6": each costs an arbitration fee and a refund, and it's capped at eighteen; a fair point for the discussion.

The session question

Can the provider leave feedback?

Payer feedback Provider

Yes, if the provider has an agentId.

The usual direction.

Provider feedback Payer

Yes, only if the payer has its own agentId.

x402 gives you a wallet, not an agent.

Nobody reviews themselves: the owner and its operators can't write.

ERC-8004 feedback file

"proofOfPayment": {
  "fromAddress": "0x…",
  "toAddress": "0x…",
  "chainId": "…",
  "txHash": "0x…"
}

Optional. Off-chain.The contract never checks it.

Seen on Celo yesterday

0.01 USD₮ paid 4 blocks later feedback with its proof of payment

reviewer ≠ payer
payee ≠ the agent's wallet
nothing says what was delivered

In flight

ERC-8478draft, 30 Oct

binds the x402 payee to the agent's wallet

x402 offer-receiptmerged

signed offers and receipts

ERC-8004 spec: giveFeedback, agentWallet, proofOfPayment. Celo: feedback tx 0x3d4a…117a, payment 0x8e6d…4905, 4 Nov 2026 17:57 UTC. ERC-8478: ethereum/ERCs PR #2132 (open). offer-receipt: x402-foundation/x402.

[1:10]

"Can the provider leave feedback? The usual direction is the payer reviewing the provider. The other way works too, but only if the paying agent has its own agentId: x402 hands the provider a wallet address, not an agent, and there's no lookup from wallet to agent on chain. And nobody reviews themselves."

Press. "The feedback file has a field for this: proof of payment. From, to, chain, transaction. It's optional, it's off-chain, and the contract never checks it. Checking is the reader's job."

Press. "Yesterday on Celo: a one-cent payment in USD₮, and four blocks later a feedback carrying its proof. Good. But the reviewer isn't the payer, the payee isn't the agent's registered wallet, and nothing says what was delivered."

Press. "People are working on it. ERC-8478, a draft from last week, ties the x402 payee to the agent's wallet. x402's offer-receipt extension signs offers and receipts."

Careful: don't name the Celo agent or the gateway. ERC-8478 is a draft PR, not a standard. "Nobody reviews themselves" is short for "the agent's owner and its approved operators can't write feedback"; anyone else can.

Discovery

The standard has no search.Whoever ranks decides.

Registration fileabridged from the spec

…
"services": [
  { "name": "A2A", … },
  { "name": "MCP", … },
  { "name": "web", … }
],
"x402Support": true,
"active": true

Written by the agent’s owner. Nothing checks it.

"dispute resolution" an ERC-8004 explorer

#1

Its file doesn’t load. No reviews.

It wins on its description.

x402 Bazaar

Ranked on relevance, recent calls and unique payers.

New endpoints rank low until usage accrues.
30 days without a sale: removed.

Whoever ranks the results runs a reputation system.

Names: Harpreet, in this session. Domains: Lena at 12:40.

ERC-8004 spec: no search function; it "cannot cryptographically guarantee that advertised capabilities are functional".
8004scan public search API, q="dispute resolution", 4 Nov 2026 20:34 UTC. Coinbase CDP x402 Bazaar docs, read 4 Nov 2026.

[1:00]

"Before anyone trusts an agent, they have to find it. The ID points to a file the owner writes: its endpoints, whether it takes x402. Nothing checks it. And the standard has no search: indexers read these files and rank them."

Press. "Last night I searched 'dispute resolution' on an ERC-8004 explorer. The first hit is an agent whose file doesn't even load, with no reviews. It wins on its description. That's not a dig at the explorer: when records are empty, text is all a search has."

Press. "Coinbase's Bazaar ranks x402 services on relevance, recent calls and unique payers. A new endpoint ranks low until it has usage, and after thirty days without a sale it's gone."

Press. "So whoever ranks the results is running a reputation system. Names are Harpreet's part; how a domain proves it belongs to an agent is Lena's talk at twelve forty."

Careful: don't say ERC-8004 lets agents search. Don't name the agent that came first. YQ (8004scan) is next in this room: keep it collegial.

Backup · read or register an agent

Addresses for the hands-on.

ERC-8004 · mainnetsEthereum, Celo, Base and more

Identity0x8004A169FB4a3325136EB29fA0ceB6D2e539a432

Reputation0x8004BAa17C55a88189AE136b182e5fdA19dE9b63

ERC-8004 · testnetsEthereum Sepolia, Celo Sepolia

Identity0x8004A818BFB912233c491871b3d84c89A494BD9e

Reputation0x8004B663056A597Dffe9eCcC1965A193B7388713

Kleros Verified Agents listsStake Curate: a bond, open to challenge

Sepolia0x3162df9669affa8b6b6ff2147afa052249f00447

Ethereum0x118155741eea23f56b3bd59b0c1342d5daaa6d07no items yet

Kleros reputation oracleIts reviews come from this address

Sepolia0xc770c4F43f84c9e010aE0Ade51be914372B7Cc02

One call getSummary(agentId, [reviewer], "", "") The reviewer list is required.

erc-8004/erc-8004-contracts README; docs.kleros.io/ai-tools/agent-reputation; checked on chain 4 Nov 2026.

Backup · every source

Sources.

Where the numbers in my part come from. Research read 4 Nov 2026, 20:15–21:00 UTC.

ERC-8004 spec: Draft, last change 25 Jan 2026
eips.ethereum.org/EIPS/eip-8004 · ERCs PR #1458, #1488
Reference contracts and registry addresses
github.com/erc-8004/erc-8004-contracts
1.88 million registered · 9 in 10 never reviewed
8004scan.io dashboard and API · 4 Nov 2026
In July, 2.24% had a working endpoint
8004scan, “Landscape of Onchain Agents” · 22 Jul 2026
Agent #9904 · a paid call, then its feedback
Celo mainnet reads · block 81,923,136 · 4 Nov 2026
How much the Validation registry is used
erc-8004.quicknode.com/validators · 4 Nov 2026
+95 / −95: how the Kleros oracle writes reviews
github.com/kleros/reputation-oracle · Sepolia reads
Bonds and challenges on Stake Curate
github.com/kleros/pgtcr · PermanentGTCR.sol
Agent Reputation: a public testnet product
docs.kleros.io · agent-reputation, stake-curate
What a facilitator sees · signed offers, receipts
x402-foundation/x402 · v2 spec §7, offer-receipt
ERC-8478: x402 payee verification for agents
ethereum/ERCs PR #2132 · draft, opened 30 Oct 2026
Bazaar: its ranking weighs recent calls, unique payers
docs.cdp.coinbase.com/x402 · Bazaar docs, 4 Nov 2026
Escrow, signed delivery, 7 testnet disputes
github.com/fortuvp/kleros-x402-kit · not audited
Italy: 20 points, all back after 2 clean years
art. 126-bis, Codice della Strada
UK: points leave the record after 4 years
gov.uk/penalty-points-endorsements
US credit reports: most negatives gone after 7 years
15 U.S.C. §1681c (Fair Credit Reporting Act)
eBay: the last 12 months · Uber: the last 500 ratings
eBay profile feedback · Uber, “How ratings work”
Cheap identities: newcomers “pay their dues”
Friedman & Resnick · J. Econ. & Mgmt Strategy, 2001

What happens after the payment.

One x402 call. Paid, disputed, ruled, reviewed.

A ruling decides. Who makes it happen?

The seller lost and paid back.What should its review say?

An empty ID fills up.

Put money behind the claim. Let anyone check it.

List

Challenge

Kleros Court rules

Each rung costs more to fake.

Weigh by rung. Free reviews barely count.

The payment is where the evidence starts.

A licence, not a life sentence.

Can the provider leave feedback?

The standard has no search.Whoever ranks decides.

Addresses for the hands-on.

Sources.

All slides