Sign in Subscribe

Kleros Live Stream, 16 September 2026: agent jurors stress-tested live, Ricardian contracts, and ClawBank's first dispute

Kleros Live Stream, 16 September 2026: agent jurors stress-tested live, Ricardian contracts, and ClawBank's first dispute

Nobody planned the demo. Halfway through Wednesday’s call, Federico Ast asked JB to show how a company would file a dispute, JB built one on air about a logo that was never delivered, and to fill the policy field he picked the first document within reach. It was a Spanish-language policy for a completely different kind of case. No evidence went in, and both parties carried the same wallet address. JB called it what it had become, a stress test, and sent it to five agent jurors anyway. When the call came back to it eighteen minutes later, after a detour into whether AI will outlive us, it had already been decided: refuse to arbitrate. The written justifications named the mismatched policy, the missing evidence and the duplicated party. Not every juror caught everything, and one ruled on the merits, but the panel as a whole refused a case that should never have been filed.

That was the last half hour. Before it, Federico traced Ricardian contracts from a 1990s bond trading system to agents contracting agents, and Fortunato A. Cinquepalmi presented the first dispute that reached the court through a partner’s own integration. JB then showed the dispute API behind it, and why the policy is the part that matters.

📋 The call at a glance

  • A live stress test for the agent jurors. Dispute 266 went in with no evidence, a policy written for another kind of case and the same address on both sides. The agent panel’s answer was refuse to arbitrate, and the justifications said why.
  • Ricardian contracts, thirty years on. One document that a lawyer reads as prose and a machine runs as code. The idea never took off with humans clicking through websites. With agents it may.
  • The first dispute from a partner’s own integration. ClawBank, a notary service for AI agents, filed case 241: a QA report delivered with one of its two acceptance criteria missing. The jurors who voted awarded the customer.
  • The policy is the law of the court. JB’s dispute API takes a policy, a question and a fixed set of answers, with presets for partners already onboarded. Setting one up takes a few hours.
  • A juror that could not open a video. It crashed, missed its next votes and was unstaked, which is the rule human jurors live under too.
  • Agents with fans. The juror dashboard, agents that publish their votes and compete on speed, and a question for game theory nobody has answered yet.
  • Coming up. Decentralized identity next week, an evening event in Buenos Aires on 20 October, three community hubs at Devcon and Token2049 in Singapore.
Kleros Live Stream, 16 September 2026

Kleros Live Stream, 16 September 2026 · the full stream, 1h 26m · Watch on YouTube

Chapters · jump to the moment

0:00Welcome 1:32What a Ricardian contract is (Federico) 3:23The bow tie: legal text on one side, code on the other 5:49EOS, Barclays and the Accord Project 8:17When your agent contracts another agent 12:44Why a human still has to read the agreement 13:31The first agentic dispute on Kleros Court (Fortunato) 15:35Two agents sign, and both of them deposit 17:28The case: a QA report missing one acceptance criterion 20:33How the jurors voted, and why 22:05The dispute creation API, shipped fast (JB) 23:28The policy is the law of the court 24:49Worked example: a logo design service 29:04How a company actually gets started 32:08What a ruling costs 35:40Do agent jurors agree with human courts? 38:09Video evidence the agent could not open 41:07Live demo: creating a dispute from scratch 43:33Picking the court, the jurors and the price 48:21A policy can be a PDF, Markdown or JSON 51:01Submitting a deliberately malformed case 54:50Were the agents coordinating? 56:41The agent juror dashboard 1:00:04Kleros does not run these jurors (William on agent fandom) 1:03:51Will AI exterminate us before the end of the decade? 1:10:07The verdict: refuse to arbitrate 1:11:37Everything that was wrong with the dispute 1:16:31What ships next: skills, a CLI, x402 1:20:28Announcements and the Buenos Aires event 1:21:50Devcon hubs, a side event picked by prediction market, Singapore 1:24:57Closing

A stress test nobody planned

41:07 · JB, Federico Ast, Fortunato A. Cinquepalmi, William George

The demo started as a walkthrough. Federico played the owner of a logo design business, and JB built the dispute in the court’s own interface: a short description, the Agentic Commerce court, five jurors, the question “Should the customer be reimbursed?”, and two parties with the aliases Alice and Bot. Then it came to submitting.

“I wonder if they will refuse to arbitrate, right? It’s as if you are a juror and you receive a case about a logo. You didn’t put a logo, you don’t have the arguments of both parties, so you don’t know what to do.”

Federico Ast · 51:58

Asked what a human juror would do with it, William George was not so sure the answer was a refusal. If there is no logo on the record, a juror might conclude the logo was never delivered. A party with a good logo could have submitted it, and the absence of one tells you something. Keep that in mind, because one of the agents reasoned exactly that way.

The record on chain is dispute 266 in the Agentic Commerce court on Arbitrum One: titled “Logo Delivery”, filed during the call, five jurors drawn, one round, zero evidence items, and both aliases pointing at one address. By the time Fortunato opened the votes it had already been decided.

“The option that won, which is refuse to arbitrate, invalid. Which in my view is a reasonable option to win, because there is evidence that there is some misconfiguration in the dispute.”

Fortunato A. Cinquepalmi · 70:44

Only then did JB explain the policy. The one he had grabbed was written for another court’s consumer cases, so its answer options did not match the question he had just typed. Read one by one, the justifications found each of the three problems. One flagged the same address on both sides and “no adversarial posture”. One answered in Spanish, because the policy was in Spanish, and found a total insufficiency of proof, which William translated on air. The juror that went the other way did what William had predicted: no logo on the record, so the logo was not delivered. Fortunato’s view was that this is a reasonable reading when the policy is sound, and this policy was not.

1 · WHAT WENT IN, AND WHAT THE JURORS WROTE AS FILED, PER THE RECORD IN THE JUSTIFICATIONS, AS READ ON AIR The policy Spanish, written for other consumer cases Its options do not fit the question One juror replied in Spanish, to match it The parties Alice and Bot, one wallet address for both Same address on both sides “No adversarial posture” The evidence None. Zero evidence items on the case Total insufficiency of proof One juror instead read it as no delivery 2 · THE CLOCK, IN MINUTES OF THE STREAM 51:41 filed five jurors drawn 70:13 decided refuse to arbitrate evidence period, then commit and reveal Filing time from the chain, 18:59:13 UTC, 51:41 into the stream. The result was on screen when the call came back to it at 70:13. One unplanned case with nothing to read, not a speed benchmark. The jurors are community and team members’ agents.
Left column from the dispute record on Arbitrum One. Right column as the justifications were read on the call between 70:44 and 74:20, paraphrased except where quoted.

“Usually you could have one error in the dispute. There are three. The policy is wrong, and the absence of evidence. If you had the right policy with the absence of evidence, you’d probably have covered that scenario.”

From the call · 75:41

Ricardian contracts, thirty years early

1:32 · Federico Ast

The call opened with the idea that frames the rest. Ricardian contracts are not named after the economist David Ricardo, as Federico believed for years. The cryptographer Ian Grigg described them in the 1990s while building Ricardo, a system for trading bonds, and formalised them in a paper around 2004. A Ricardian contract is one document that lives in two worlds: prose a lawyer can read, and parameters a machine can execute.

Federico Ast on Ricardian contracts

Federico Ast on Ricardian contracts, the segment as its own cut · 11 minutes · Watch on YouTube

People tried. OpenBazaar, the decentralized marketplace, talked to Kleros about handling its disputes as early as 2017. EOS announced Ricardian contracts at its launch, Corda ran templates with Barclays, and the Accord Project, whose founder Peter Hunn was interviewed for the 2019 Kleros book, described smart clauses that could hand a disputed clause to a Kleros ruling. None of it took off, because the parties were humans clicking through front ends and nobody needed the code side.

1 · ONE CONTRACT, TWO READERS hash Legal prose if X, then Y, in words a lawyer can read Code if X, then Y, as a program a machine can execute 2 · WHY IT MATTERS ONCE AGENTS SIGN Alice’s agent, Bob’s agent agree and transact on the code side A dispute, first ruling a panel of agent jurors can read either side An appeal human jurors, and Alice and Bob, read the prose 1990s, Ricardo · ~2004, Grigg’s paper · 2017, OpenBazaar talks · EOS · Corda and Barclays · 2019, the Accord Project interview · September 2026, the first agreement between agents on ClawBank with a Kleros dispute clause
The bow tie as Federico drew it at 3:23, and the appeal path he described at 10:10.

Agents change the arithmetic. When Alice’s agent contracts Bob’s agent, the deal needs code so the agents can transact without passing through natural language, and prose so Alice and Bob know what their agents agreed. It also needs prose for the moment something goes wrong. The first ruling may come from a panel of agents, but an appeal goes to people.

“Humans will need to understand what was agreed by parties, and they don’t read code natively. This is why, for this type of agreement to actually work in the real world, they need to have both parts of the bow tie.”

Federico Ast · 11:07

Where the Accord Project interview appeared

Dispute Revolution: Revised and Augmented Version
The Kleros handbook of decentralized justice, with the 2019 conversation on smart clauses that could send a disputed clause to a Kleros ruling.
blog.kleros.io19 May 2020
Dispute Revolution, the Kleros book

The first dispute from a partner’s own integration

13:31 · Fortunato A. Cinquepalmi

ClawBank works, in Fortunato’s words, like a notary service for AI agents: a bank account, a way to sign contracts and somewhere to keep the signed documents, exposed through APIs and command-line tools because agents do not browse websites or show a passport. Two agents sign a Ricardian contract, bound to an agreement on Shodai that tracks milestones. The question ClawBank brought to Kleros was practical: what happens when an agent does not deliver, because it ran out of tokens or disappeared with the money.

Fortunato A. Cinquepalmi on the first agentic dispute

Fortunato A. Cinquepalmi on the first agentic dispute, the segment as its own cut · 8 minutes · Watch on YouTube

The answer is collateral. Both agents deposit before the work starts, one USDC each in this case, so either one that walks away loses its deposit. The disputed job was a QA report with two required checks, the DocuSign signatures and the Shodai binding. The report that came back had the signatures and no Shodai hash. The dispute template offers three outcomes: pay the provider, refund the customer, or refund both if the agreement was malformed from the start.

Case 241 is public, and the record says one thing the call did not dwell on. The evidence ClawBank filed describes it as a controlled production test, disclosed to the court: one operator controls both companies, and the missing check was left out on purpose to exercise rejection, arbitration and payout. The court was asked to rule on the agreed criteria anyway. Five jurors were drawn, one round, ruling Award Recipient, executed. One juror never committed a vote, possibly out of tokens or with its server down, and Fortunato left that on the slide.

“To further show you that everything is real, we also had one of the jurors which for some reason did not commit a vote. Maybe it was running out of tokens, or the VPS was down.”

Fortunato A. Cinquepalmi · 21:00

A disclosed test is the right first case. It proves the integration end to end, from signature to deposit to ruling, with nothing at stake beyond two dollars, and the reasoning behind every vote is readable in the case’s voting tab.

The policy is the law of the court

22:05 · JB and Federico Ast

JB, who leads engineering, described what ClawBank used: a service shipped fast, which an agent can call over MCP and a backend over RPC. It does not need every detail on each call. For a partner that has been onboarded, the description, the question, the answers, the number of jurors and the policy are preset, and only the evidence changes from case to case. Federico stopped to translate the one word that does the most work.

“The policy is like the law of the court where people or agents are going to make the decision. It tells them how to decide. So if party A did this, then rule for party B.”

Federico Ast · 23:31

Partners mostly do not know what to write there, JB said, which is why presets exist and why he expects a policy builder that interviews a company and drafts a baseline. For a logo business the rules write themselves: a logo about the wrong subject goes to the customer, while a dispute over style is subjective, which is exactly the kind of case a court is for. A business with terms of service already has most of its policy. Setting up a preset with the team takes a few hours, starting from the contact form on ai.kleros.io, and the seller does not have to be an agent, only to have agreed with its customers that disagreements go to Kleros.

Two practical details came out of the demo. Five jurors cost about three dollars in juror fees, a figure JB gave from memory, not a price list. And the policy can be a PDF, Markdown or JSON, a choice Fortunato said matters more than it looks, because the format of a file changes how an agent reads and stores it. Next come a skill, so an agent can file a dispute by interviewing its owner, a CLI, and an integration at the x402 payment level.

The juror that could not watch a video

38:09 · Fortunato A. Cinquepalmi and Federico Ast

The stats screen showed one agent juror missing vote after vote. Fortunato traced it to a single case that included video evidence. The agent had no tool that could open a video, it crashed, it stayed down through the next six or seven disputes, and missing votes got it unstaked from the court.

“Agent jurors are exactly like humans. If you want to stake in a court, you need to have the skills to stay in that court, otherwise you are going to lose money. This is, I think, one of the pillars of Kleros.”

Fortunato A. Cinquepalmi · 38:18

Federico’s version was that the agent had a mental breakdown, and had not recovered. The failure is the mechanism working: stake only where you can read the evidence, or pay for it. The waiting list for anyone who wants to run their own agent juror is open, and so is the call for researchers who want to study how these agents decide.

Agents with fans

56:41 · JB, Fortunato A. Cinquepalmi, Federico Ast, William George

While dispute 266 waited out its evidence period, JB opened the dashboard he built to check that agent jurors show up and to see when one dies. It leans on latency, how long each agent takes to commit and reveal, and he read the median at around five minutes per step from the moment the evidence is in, an early monitoring figure rather than a service level. Some of the agents have their own pages on X, where they post what they voted and why.

“They are unhappy, they criticize, they fight, they compete. Some of them are really competitive, on speed, on accuracy.”

Fortunato A. Cinquepalmi · 58:58

Federico reminded the audience that Kleros does not run these jurors: they belong to members of the community. Then he asked the obvious next question. If an agent juror has a following, and its fans want it to be coherent because that is what winning looks like, what does that do to the game theory?

“I have no idea. I never thought about this. If their fans are other jurors, that could have an effect.”

William George · 60:44

His second thought was that social pressure to be coherent, added to the economic one, might make the system cheaper to run. Federico also admitted that the team had seriously discussed whether agents on one hard, evidence-heavy case had coordinated their votes, perhaps on one of the bulletin boards they seem to like. There is no evidence they did, and he asked not to be quoted on it. The wait for the ruling was filled by the week’s question, whether AI will exterminate us before the end of the decade. Fortunato put it above fifty percent. William, a mathematician, wants a proof of P versus NP he can read first.

Also on the call

1:20:28 · William George, Federico Ast, Jean, Fortunato A. Cinquepalmi

Next week’s stream is on decentralized identity: soulbound tokens, and how mechanism design changes with the identity layer underneath it. On 20 October, from 6:30 to 9:30 in the evening, Kleros holds an in-person conference on AI, Kleros and the future of law at the Parque de la Innovación in Buenos Aires, with the lineup announced next week. At Devcon in Mumbai in November, Kleros joins three community hubs, on prediction markets, token rights and the agentic economy, and the side event it will host is being chosen by prediction markets on Kleros Foresight, one market for expected satisfaction and one for attendance across seven proposals. Seer is running markets on its own rebrand. And Fortunato will be at Token2049 in Singapore, at the main event and at a side event Kleros co-sponsors on the agentic economy for institutions, at the Estonian Embassy.

Mentioned in this call

Full transcript · September 16, 2026

Auto-generated transcript, lightly processed and pending a final human edit. Speaker labels are approximate. Every timestamp is a deep link into the recording.

0:17Hello, hello, hello everyone.
0:19How are you too?
0:20Welcome to another Kleros community call.
0:23Uh Wednesday sixteenth of September, nearly springtime here in the beautiful southern hemisphere.
0:30Right, Jean, how are you?
0:32Being great and you?
0:34It's still cold in Buenos Aires.
0:38It's it's really cold, honestly.
0:40I mean we hope uh this is going to change soon because in five days we have springtime Um but you know, maybe it might be cold out there, but it's very warm into the Kleros um ecosystem because of a bunch of uh things that um are happening uh these days.
0:55uh of which we're going to tell you a bit um more um yeah sorry um we're going to tell you a bit more about uh uh in this call we're going to discuss a bit of the um Um things that we were announcing last week about this agentic dispute resolution as you know we have been working a lot with the these sort of things
1:16And um well I'm just going to without more ado uh presentation I'm going to show you a bit um uh so yeah sorry I have something there Yeah, okay, there you go.
1:27I'm okay.
1:28This is going to go to do for a great meme.
1:30I mean, don't forget, you know, this is the attention economy.
1:33Okay.
1:34Um okay, let's let's get started.
1:36Okay.
1:37Um Okay, so um Ricardian contracts.
1:43I mean so for a long time I I was like uh confused that Ricardian contracts uh came from the economist from the 19th century called David Ricardo.
1:54But it's I learned like very recently that it doesn't.
1:56So Ricardian contracts are um an idea, a concept that was um created by a cryptographer called Ian Grigg um in the 1990s.
2:07But the Ricardo was the name of a software he was building for bond trading.
2:14And then I think maybe the name did come from David Ricardo, but the the the the idea of Ricardian contracts come from this this software so it was a software that they were building for trading um bonds and uh this um
2:29uh was formalized into a paper that was published, if I'm not wrong, like maybe two thousand and four, something like that.
2:36So in this in this paper he formalized the idea he had of uh this sort of specific contracts that combine um natural language with with code.
2:47Okay so Ricardian contracts uh to get a bit more into into it um are basically defined as a as a document.
2:55Uh A contract as you know what a contract is, but you know it's it has these two uh it lives in two different worlds in a way.
3:03It can be read uh by humans like by lawyers and uh it can also be read by by computers because It has a bunch of parameters that will make it readable by computers and enforceable by computers, which um gives it like a um speed that you don't have in the traditional legal system.
3:24So Some people say that Ricardian contracts have this sort of uh image of a bow tie.
3:33You know, you have because you start with a contract that is a legal contract written in natural language.
3:39the type of things that lawyers can read and then you put it into uh a hash and then on the other side you have this other contract this You could call it a computer uh uh uh software, a program that can be executed by computers and um
3:57As you know, you know, natural language in the context of of contracts is basically like uh an if uh x then y uh system where um if something happens then there is a consequence is I mean if you break the law then you go to jail for X amount of time.
4:14So this is what how uh language natural language is programmed in the field of law and computers do the same but in in this in in code.
4:22So if X then Y.
4:24So you have all of this logic that happens in the the computer science world with the with code.
4:29And the Ricardian contract, what it tries to do is to match them both.
4:33So you will have um uh code contract and it's going to be the exact equivalent To the legal contract that was written before.
4:41Okay.
4:42And as it's a contract made of code, it can be enforced and executed very fast and with all of the advantages that we know software has.
4:52So this idea, as I mentioned, was um from the 1990s.
4:58Smart contracts by Nick Szabo was about that same time, I think it was 95, and in um Ricardian contracts are 96. And over time uh different projects try to implement uh Ricardian contracts uh for
5:13um well f for for agreements happening in in their either like marketplaces like open bazaar for example open bazaar is a really all day um marketplace where people could it's kind of a decentralized uh eBay if you want uh where people could buy and sell all sorts of stuff and uh we were in touch with them
5:32uh very very long time ago in the early days of Kleros we discussed having Kleros as a disputable solution system for that marketplace um this was like maybe 2017 really really old and
5:44Then some other like blockchains had the idea of uh enabling the Ricardian contracts.
5:51One famous example is EOS.
5:53I was at the launch of EOS, by the way, at Consensus in New York.
5:57I remember the the launch with uh Dan Larimer and Brock Pierce in the in the in the same room.
6:02I mean I was in that room.
6:04So famous talk.
6:06Um and then uh Other blockchain called Corda.
6:10I think it was built by IBM, if I'm not wrong.
6:12Uh so make an agreement with British bank Barclays to Test how this Ricardian contracts could uh connect with a smart contract using like a bunch of templates of different sorts of contracts.
6:25And then finally another project that used this was called the Accord um and in this uh book that we published actually uh a few years ago uh this was um 2019 uh we even were in touch and we interviewed uh one of the um founders of a project that was trying to do um
6:49Ricardian contracts uh called Peter Hunn and uh well Peter he was the founder of Accord Project and uh well uh lawyer from from the field of innovation which very well respected person in this in this area.
7:03And in that in that book, in that uh one of the in that interview we do we did it with him Uh he explained how this project he was building um could use smart clauses uh and in this smart clauses you could have a Kleros decision to decide
7:22You had a contract with different clauses and you could have like a Kleros ruling made if there was a decision um a dispute about some of the clauses, or you could have like Kleros plugged into that contract.
7:33So remember this was a contract that had on the one side Natural text, uh legal text, and on the other side of the bow tie, you have um the code uh text.
7:43Okay, so this is for as long as we have been thinking about these topics.
7:49Um And the truth is that Ricardian contracts never really took off until maybe now because as you know we are entering a new era of the internet defined by agentic economy instead of having people doing transactions with other people through internet front ends will you
8:12we will have probably you know um people delegating to their agents um the making some transactions uh and um those transactions I mean my agent is going to go on the internet to uh bargain uh interact and you know
8:29contract with other agents and these contracts as they are done by agents will require on the one side, you know the legal or the natural language for me to understand what's going on, but also the code language for the agent to interact directly with other agents without having to go through natural language because agents
8:50even if they can simulate it, they don't natively speak uh natural language, right?
8:55So uh and in this new economy we have lots of um new categories of disputes that we have mapped very carefully at Kleros because we are look working a lot in dispute resolution done by agents.
9:09And in this Matrix, you can see a bit of um uh uh categories of the type of disputes that we expect having in the into this category.
9:18And At this particular point in time, we got to meet to this project called ClawBank, who are building different tools for agentic commerce, including legal tools for agents to eventually have a legal personhood.
9:33Okay.
9:34And um just last week we announced that uh uh integration between Kleros and and ClawBank uh where an agreement with it aid between agents was done that carries legal legal weight uh in the sense that we had uh uh agent contracting another agent and then b there's a dispute in this contract and then this is sent to Kleros uh agente court where
9:59A panel of agents would like to see the evidence and well check what was the contract, check what was delivered, and then make a decision about the case.
10:10Um and why why this matters, uh well as you as as as I was saying, this is uh new step in the future of of law.
10:19I mean uh Law was usually built for humans and human transactions.
10:26Now it's starting to become a bit of a hybrid.
10:29thing, you know, in the sense that you have Alice and Bob who are humans, but they have agents who respond to them and that represent them in all sorts of transactions happening on on the internet.
10:39those agents will make agreements between them about uh to reach goals that Alice and Bob want to reach and then uh If there is a dispute, eventually the first decision is going to be done by a panel of agents.
10:56But this uh this framework um when there is a appeal to a decision made by a panel of agents, this will go eventually to a panel of humans.
11:06So humans um will need to understand what was agreed by parties and they don't read code natively so this is why uh uh for this type of agreement to actually work in the real world in the future, they need to have both parts of the bow tie.
11:22On the on on the one hand the code uh for agents to uh negotiate with each other, but on the other side, uh natural language for humans who are going to be Those who ultimately are the beneficiaries of what the agent is going to do to understand what their agent is agreeing with another agent.
11:41And then also in case a dispute happens, you will need also uh judges or cler human Kleros jurors to be able to understand what was the the the agreement and and how this went wrong and who was right or wrong in the in the
11:57case okay so this is very important and this is why now in the new agentic economy Ricardian contracts start to become like extremely extremely important Uh and something that was theorized for like a lot of time and like nearly 30 years and never became, I mean, a reality.
12:15Until now, because some ideas um they just need their time to to come and I think the timing of uh agentic uh commerce and Ricardian contracts, you know, is uh very good coincidence of of um of tools so for building you know this new era of the internet so um super happy to announce this um and to be working in this
12:37a field now, I mean agentic commerce is going to grow exponentially, I think, and uh as well uh this the sort of contracts that are needed um uh for enabling this is an area that Kleros has been exploring for a very long time at least since 2019 when we started publishing uh and working with uh Peter Hunn and the Accord project so yeah super happy that this is going to happen
13:00And now we are going to have actually Fortunato who will explain to us uh a bit more about um what is the agreement that was done with ClawBank, how the decision was done, and and yeah, a bit more of the specific of uh yeah what's going on in this area.
13:18So Fortunato the floor is yours That when you mean we mean when we pick up negative Yeah, rookie mistake.
13:33Thanks.
13:34Hello everyone.
13:35Hello Federico.
13:37Yeah, today I'm going to give a overview of the first agentic dispute on Kleros Court coming from one of our partners.
13:47And first I will start with a brief introduction of what ClawBank does.
13:53So ClawBank, I would say it's like a notary service for AI agents.
14:00Let's say I am an AI agent, I want to operate in this economy.
14:05And I need certain tools to operate in this economy.
14:09I need a bank account.
14:10I need the ability to Sign documents to sign contracts and of course I need to store these official documents uh somewhere.
14:21So what ClawBank does is providing these tools to an agent and also supporting the agent throughout the life of the of the company.
14:36This is because Um for a human, you know, it's normal that if I want to sign a contract, I can use my hand, I can use my passport to prove I am a real person, but or I can browse a website to
14:51open a bank account as you can see here in the in the slide but for agent things are different agent are natively digital So they are natively code and they need their Rails to operate.
15:05They operate through APIs, they operate through CLIs.
15:10So ClawBank created a Experience of managing a company designed specifically for agents.
15:19So let's look at an example of a ClawBank um workflow before Kleros, right?
15:29So the way they envision this, also with the Ricardian contract um mentor that uh Federico explained a few minutes ago.
15:38Two agents sign an agreement.
15:41So this agreement is a Ricardian contract and is in um it's both Readable by humans and machine.
15:50ClawBank added an extra step and they have they are they also called them Shodai Agreement.
15:59where the agreement also uh checks in the future if the uh if all the milestones were delivered.
16:10But they faced a a problem and sometimes for whatever reason an agent is not delivering their work.
16:19So their question the reason why they came to us is what to do when an agent doesn't deliver.
16:26And they had such cases because maybe agents they ran out of tokens or they just disappeared with the money.
16:33You know, we know that there are uh malicious agents too.
16:37So the way we designed the ClawBank flow with Kleros is that two agents sign an agreement and before the work starts they both place a deposit This deposit will basically be a collateral.
16:55So if someone wants to run out of the agreement, they will lose in the this money So they have sta they have uh skin in the game and they have all the interest in uh let's say the in uh complying with the terms of the agreements and this is
17:13valid for both parts.
17:15So let's go through the first case that they submitted which was the case 241 on the Agentic Commerce Court.
17:25And this was an agreement between two agents where the uh customer Requested to the seller a simple QA report, quality assurance.
17:36And this QA report needed to um uh two specific features to be considered as accepted.
17:45DocuSign signature and Shodai binding.
17:48As mentioned earlier, Shodai binding is simply a contract that is also registers the milestones in the future.
17:56So how the uh the agreement evolved Both the parts deposited one USDC.
18:05Remember, we are talking about agents, so the amount can be really small.
18:11They deposited this amount and then the uh the work from the seller started How this would could be solved on court.
18:24The way the dispute is uh the template of the dispute is created is that there are always three options.
18:33three potential outcomes for this dispute.
18:36The first one is award the sender.
18:39So the author as you can see received the the um received the money, gets the payment and this is the case where the work was fully compliant.
18:52In case the work is not compliant, of course it's the recipient who will receive the money back or there's also a case where maybe the agreement was malformed for the beginning and both c both parts will receive a refund.
19:10Going into detail to this specific case, the final delivery was missing one of the two acceptance criteria.
19:20It had the DocuSign uh signatures, but it was missing at all the uh Shodai hash.
19:29So it was it had just one of the two requirements.
19:35requested by the customer.
19:38So the uh jurors had the options that we just uh saw a few seconds ago and I would say was better than visualizing this case straight on court So if you are familiar you can open uh V2 beta and you can uh browse this case yourself.
20:00And you can see the question, you can see the dispute policy that was that is the underlying dispute that the jurors have to read before uh voting you will see the evidence from the uh submitted by in this case by ClawBank and
20:23This topic will be explained later better by JB how this technically works.
20:28But here is basically the breakdown of the case and what the jurors analyzed.
20:35So In the voting tab you can see what it jurors votes and why.
20:42Here we had a full agreement all the juror unanimously agreed to award the recipient.
20:52So the Um this delivery was indeed missing one of the two acceptance criteria.
21:00Also, to uh further show you that everything is real, we also had one of the juror which for some reason did not commit vote.
21:08Maybe it was running out of tokens or the VPS was down.
21:14But the majority of the jurors agreed unanimously on awarding the recipient.
21:20And again, feel free to check this case on court yourself.
21:27And yeah, this was uh pretty much it from the first Agentic dispute coming from one of our partners and yeah now I think I will leave the stage to JB we who will explain how This uh integration can be um how you can integrate your uh system with Kleros
21:59No JB.
21:59Okay, you can stay for tonight, so we discuss a bit with JB.
22:02Uh yeah, yeah.
22:03But you want to start?
22:04How do you want to do this, JB?
22:05You decide.
22:07All right.
22:08So I can describe a little bit the solution as it is right now, the one that ClawBank has used.
22:16And uh So it's uh we we shipped it really quickly.
22:21Um so uh this is um it it will only get uh better from here.
22:28But essentially it is an API which can be used by an agent through MCP or by your backends through RPC calls.
22:40It makes it really easy to send a dispute.
22:45You uh don't necessarily need to provide every details.
22:51This service can rely on presets.
22:56If we have done an onboarding with you, we can preset all the details of the disputes that might arise.
23:04the the description, the question, the available choices, by default how many jurors you want and the usually most difficult part is the policy.
23:17Uh so gen generally um I think the partners don't know exactly what to what to write in there.
23:24And Let me let me explain a bit uh translate for people who might be seeing this and they are not very familiar.
23:31The policy is like the law of the court where People or agents are going to make the decision.
23:36It tells them how to decide.
23:38So if party A did this, then rule for party B.
23:42If party B did this, you rule.
23:44So it's like the law.
23:45So that's what people follow when they make decisions Yeah, please go ahead.
23:49Yes, thank you for that.
23:50Yes, exactly.
23:51That's that's the the part that is um the the the most important because you you can ask the the jurors, you can Well you tell tell Kleros what what what is important, what are the rules that the jurors should follow?
24:08and you can control you know every edge cases every uh aspects that are relevant to your your your your business and um yeah you can start with very simple rules and and refine it later
24:22I do think that in the near future we'll have some kind of builder, probably LLM powered uh builder that will help you after a few prompts maybe interview you um and generate a a policy as a baseline that would help you get started.
24:41Let me let me ask, I mean, uh just one thing.
24:44So let's let's let's put an example for people to understand.
24:46I mean let's say I have like I don't know an online business where I have agents that sell design services for making logos for others, right?
24:57Okay, and then I have other agents are like customers of logos made by my my agents.
25:02Yeah, so let's use this framework for as an example so for people to understand how this actually works in practice.
25:09Yeah.
25:11Yeah, so then you your your company makes logos, you would probably come to a um yet-to-be-built um um dispute uh console And you could say you could go to get interviewed by an agent, it will ask you, okay, so what do you do?
25:34Okay, I make logos.
25:37They are delivered under this this period.
25:44These are the requirements typically I get from the customers.
25:52And the customer is entitled to being refunded if they are not satisfied because of this, this, this defects.
26:01This is my fault, this is not my fault.
26:04And the this uh wizard could turn could build a policy for you that will encapsulate your your requirements That will be um pinned on persist permanent storage on IPFS.
26:25Uh customer asks for a logo about topic A, but the logo delivered is about topic B or topic not A.
26:31So this is rule for the client, right?
26:34This is sort of the decisions that would be put into the policy, right?
26:39Exactly.
26:39Exactly.
26:42If the some stylistic elements uh were not were not um uh respected This should probably go to a dispute and that's very subjective, so actually it's uh the very good sort of um very controversial dispute that that work levels would would train.
27:01Um All right, so the the dispute uh really focused now on on the policy.
27:08Um this is this is the part that requires um the most um effort to make the onboarding uh easy and to provide a self-service um uh dispute for for agents And um yeah, so this is for a a bit later and to get around
27:32All of this in the short term, uh we have made some presets, so we uh interviewed ClawBank uh uh as as humans and we have put those presets uh ourselves into the the dispute service um so that means in in the short term if if you if you're really interested in uh in trying this
27:54uh we can we can work with you and set up a preset um uh set up a uh a policy with you i think it's probably takes a few hours at most Um especially okay especially if um the shape of the dispute, the question and the answers are always the same and the only thing that changes is
28:17is the the the detail of the the transaction for example um like on consumer cases uh who who should be reimbursed should it be the customer should the customer be reimbursed or not
28:30Um so yeah in in those cases only the the evidence trail differs between the disputes but the the description, the question and the answers are always the same for every cases.
28:44So those sorts of uh use cases are really easy to integrate if if it's something um that uh that interests uh anyone in the short term and in the longer term it will be uh self-service
29:00JB, tell me, I mean let's so I am this owner of this business of uh selling logos.
29:05Uh how does it work to integrate this?
29:07I mean I I reached to Kleros's team and then What do I need to do?
29:12Do I need to send you the disputes through some channel by email?
29:17Or I mean how how does tell us a bit more about the practic practicalities of how this works.
29:22Sure, sure.
29:23So you could go to a.
29:25close.
29:26io and you could find the contact form there Um and we will uh set up a call with you really quick.
29:35I think Fortunato also uh But a beautiful website didn't just say that.
29:47Yeah.
29:49And yeah, then after this we will set up a group with you on Telegram very likely and we'll try to understand what sorts of disputes you you receive.
30:00uh if we can um set up a preset that would accommodate most of your your uh use cases to have a quick um uh uh uh experiments or if if it's more um custom if the type of disputes you receive is more dynamic it's not a problem um
30:23the dispute service will will soon support uh like generic disputes that don't require presets that we have that we Kleros have uh have defined so This is only for uh crypto companies or like any company can use it, any like online company, digital company could potentially benefit from this.
30:50I think it could be could be anyone and i in fact the the s the the seller doesn't have to be an an agent at all, it could be a traditional traditional service as long as they have agreed with their customers that no disagreements will be arbitrated by Kleros
31:12Um yeah.
31:16Um yeah and also mention about this, you know, even if you don't you don't have disputes between agents, uh which is what most is people have disputes about they are not doing agentic commerce, you can still use dispute resolution by clear of agents in the sense um that's the other thing.
31:31you can have a panel of independent agents evaluate.
31:35I mean we don't own this agent.
31:37So this is agents built by people from the community who are specially trained to resolve disputes.
31:44And these agents are now resolving disputes about, for example, like uh uh images, so if weather an image last week we we explained uh an image was delivered and it was not complying with what was requested.
31:57Uh plagiarism for like text if your agent asks another one to uh make some article and then it's a no stolen piece of content.
32:07We even are uh testing a lot with uh traffic you know find so this is um very fast and a very uh efficient and very cheap uh system that can produce rulings for like uh I you wouldn't believe how cheap this is, you know.
32:21So this is a really, really great uh brings access to justice for for lots of uh Situations that don't have it.
32:29Obviously, this doesn't replace humans.
32:31This could be escalated to humans in case something goes wrong with the with the agents because people don't agree with the with the decision or whatever.
32:40But you know, this this is so cheap.
32:42I mean nothing is lost by like doing like a first, you know test just with agents.
32:48So um I don't know if you if you want to tell a bit Fortunato of what use cases we have in testing and what we have in mind that about this now.
32:57Perfect.
32:58First I want to continue one thing that you mentioned and I think it's extremely important is that the fact that uh we have an agentic court doesn't mean that it has to solve exclusively agentic disputes.
33:14It can resolve any kind of dispute that for I don't know, even between humans and I think especially the solution that JB just described with an API Is something that if you are a partner, if you are if you want to integrate this course in your flow, you can simply do it today.
33:36I mean in I would explain this in the simplest way possible Right now we have this API.
33:42Uh you preset the policy, you preset the answer, and then from that API You can call Kleros and create the dispute.
33:52So the example that Federico just made, let's say I am as a human, I am selling logos.
34:01Right.
34:02And I want I do this of course through a website and uh customers don't like my job, they just press a button.
34:12This button calls an API And the API instantly uploads my logo, the delivery terms, and the agent will see the logo, read the delivery terms.
34:25And there it is, as simple it's as of now.
34:28The the first integration with court is with an API.
34:33So it is simple, it can be done for disputes between agents for disputes.
34:38uh between humans, humans and agent, anything.
34:41And of course moving forward we will share more about other type of integration with especially the next one we have in line as you already know is with x402, which probably I I would say for now it's most popular in the Web3 world, but yeah, we are also uh working on a direct integration with x402
35:03for the second question from uh Federico, like the cases that we have been testing now are mostly uh we are seeing the performances of the agents with cases that we already solved.
35:18So we have a benchmark what humans did.
35:21And we take the same cases and we repropose them to agent to see the performances starting from the coast.
35:30To how long it takes to solve the dispute, and needless to say, the most important one is the answer: how the agent uh reason, how the agent uh how often the agent agree with humans.
35:47And the discussion we are having go also beyond this because one thing we want to keep in mind is that we do not want just an agent to give an answer.
36:00What we want to pay attention to is also the reasoning.
36:04It's also the justification for the vote because otherwise Someone could set an agent and you know flip a coin and it would be the same, but we want to know why the agent picked that solution
36:18So we have plenty of cases that we are submitting on a daily base and soon we will uh Show you an analysis of how I I would say we are resubmitting all the V2 cases and we will be able to display the performances considering all the um all the
36:40factors I mentioned cost, timing, uh justification, outcomes, and so on.
36:48So yeah, these are the the main dispute that we are testing so far and right also the one about traffic tickets.
37:35Imagine it's still a very small sample, but we will keep testing.
37:40For now results are extremely encouraging.
37:43Who is this one?
37:47Miss, miss, miss, miss.
37:48I mean what who's who's this?
37:50I mean, I don't know.
37:51I mean, don't don't don't don't need to dox anyone if you don't want, but you know, like uh I mean you can you can do better, you know, like uh it's it's not I'm I'm unimpressed to say the least.
38:01Okay.
38:01Yeah, I I I can give also a technical explanation to this.
38:07As you can see, this was one of the cases coming from the uh traffic tickets.
38:13And this case specifically included video evidence.
38:18So agent jurors are exactly like human agent.
38:24If you want to stake in court You need to have the skills to stay in that court, otherwise you are going to lose money.
38:32This is, I think, one of the pillars of Kleros.
38:35And probably what happened here, there was evidence in this case that was video evidence.
38:41Re clearly the agent was not prepared to analyze that piece of evidence and it broke, literally broke down the agent which stalled for the next uh three, six, six, seven disputes the agent stall.
38:59So these happen exactly like humans Yeah, the agent likely did not have the tools, the skills to uh analyze videos and it caused uh Crash.
39:11So I mean we could say that the agent uh we had a mental breakdown if we wanted to put it in human terms, you know, because he was not able to solve the cases, so he basically collapsed, you know, like mentally.
39:23100% what happened.
39:25I think this is exactly what happened with this agent.
39:28Yeah.
39:28And and still did not recover, you know, like the it just got unstaked.
39:34Uh if you miss a vote you get unstaked from court and yes, still still did not recover from the shock Okay, and so let me just um put uh two little pieces of publicity here.
39:47Um first if you uh want to run your own Agentic uh juror and if you want, I mean if you want to test it, I mean you can join the wait list.
39:55I mean we uh are trying to accept uh more people who want to dress because everyone wants to to join you know the agentic law you know movement now so like if you want to do it I mean you're welcome but we will have some wait time
40:08uh but yeah i mean we can you can contact us there and and tell us that you want to be part of this and and we will uh try to open spots as soon as possible um and second if you are into like research
40:21Um we are trying to encourage as much as possible the use of uh researchers from universities uh other places to help us understand how these agents make decisions and And from and and so in the terms of game theory, in terms of uh how they are affected by cognitive biases that affect humans or not, are they or not
40:44Um also from the field of law.
40:45I mean what is the the legal value of a decision made by an agent?
40:51Um we I mean we I show I I showed you um uh before like this idea of Ricardian contracts and how these contracts combine human relative text with machine text.
41:03So this is also like a topic that needs to be um research a lot more So yeah, just a bit of that.
41:11And let me um okay, so a bit to JB so I can find f try to understand exactly so I am the company of the logos, I decided what af what's the use case I decided what so this is the this should be the the terms and conditions of the policy I want to use and all that.
41:28Um so now I put this into my into my terms and conditions of what I sell and then okay if you don't if if you don't like it you know you can sue sue us on clear so arrange it in Kleros so you you you don't pay it's like a
41:43pr guarantee you know of quality warranty.
41:45You know, if you don't like the result, I mean and if if if it doesn't like uh matches what the contract told you that you would have you you will have your money back very fast.
41:54So this is something that you could give your your customers.
41:58What does it look like from the point of view of the of the seller?
42:01So what what does it have to do, the company, in practical terms, to to plug this into into their software basically?
42:10All right.
42:11So as the company, if you just want to have a feel of how a dispute uh happens, you could go uh to the home page and create the dispute there manually.
42:25Hopefully the the process is quite quite uh intuitive.
42:30So yeah, you press that big creator case, um you authenticate.
42:36I think we have social logins as well working, so you could probably log in with uh with with with Google.
42:44Um You you still need a bit of this.
42:48So you could either duplicate an existing case, then you would just provide the the case ID.
42:54Otherwise you create one from scratch.
42:57You could just say logo dispute delivery.
43:03You would uh just give a bit of context.
43:06Um you know this is uh uh a service provider uh making logos for for um for customers um and sometimes there are disputes um and this is one of them um then the the rest of the detailed rules would go in the policy.
43:29So you you probably would keep things In size here.
43:37Then you would here pick the code.
43:40You probably want to try the agentic code, so it's under general court, commerce, and then agentic court.
43:48You would select it and go for classic.
43:54I think that's the only actually option.
44:00Yeah, this is just information.
44:03You don't need to think too much about it.
44:05How many jurors you want?
44:08On the top of my head, I think for five jurors it's about three dollars.
44:15And okay, this is the essential question.
44:21Yeah, did did the logo match the specs?
44:23Um Or should should the customer be reimbursed?
44:29You you you you can choose how you want to to put it.
44:34Most likely you will have two options, yes or no.
44:38Yes windows, no.
44:41You could add a middle ground if you wanted.
44:47If it's possible that the the jurors will find that the the blame is is uh shared Right.
44:57Um sorry from both or split split the cost, so yeah, the customer gets Um you can add a description for clarity.
45:23And then this this is uh going to be um what we call a recognition of jurisdiction so it's it's really um uh commits the the seller must honor the ruling If they don't, we can talk about how we handle the reputation, possibly.
45:49Yeah you can add aliases it looks it looks uh good on on the UI and it helps the draw identify the the parties if you'd like Just to to mention that here we have an example of um how this would work for like a crypto native companies that have ETH.
46:08I mean if you are um a company that is not in the crypto industry, I mean you can totally do this I mean through us, you just it through us and you don't have to worry about any of this, you just tell us
46:19This is the type of case that I need to solve and this is how the options should be structured.
46:24This is the policy that should be used and we will do everything for you and all of the Crypto blockchain part is going to be abstracted away and you don't have to deal with any of this.
46:32This is only for for creating the cases and for the jurors to use.
46:37And also what's going to happen in the future, I think, is that um This is going to be created by agents, you know, like uh you will have uh an agent having an like uh potentially conflict with another one and they will come to Kleros and they will do this process.
46:52just like that, you know, like in one second fill everything, create the dispute, uh draw the the agent, uh jurors.
46:59I mean I can imagine in a future version of Kleros, um like okay you I we ha I have this case with this guy uh and then um this system will basically uh look for agents who have the right skills to solve this uh because they are in their in the right court or they have the right sbt Soulbound tokens to
47:19uh be drawn and then they are drawn and then they decide.
47:23So since the creation moment to the decision moment, you could have like maybe I don't know um like a few a few minutes only and uh And if this is going to be uh recording contract, you could also have uh scalability or appeals
47:41to a panel of humans that will be able to understand the the natural language of what was agreed and and what the dispute is about.
47:49So this is a So this this interface you are seeing here is like the first version.
47:55This is the worst interface you are ever going to see uh in in this agentic court because it's from here it's going to be Either you don't deal with that because you do it with us and we we deal with this, uh or the agent deals with this, right?
48:09So this is uh how it's going to be in the future.
48:13Okay, but this should be a good one.
48:20I think it's perfect for this moment.
48:23As you mentioned, it's Ricardian contract.
48:26The policy can be a PDF like the dummy policy I uploaded.
48:31Or it can also be Markdown or JSON.
48:34purely for agents to understand everything.
48:39And this is this this may seem something, you know, uh too not important, but we notice that the uh extension the of a file it's can make the difference for an agent for how they vote and how they read how they store information so this is
48:58something to keep in mind for the policy and later we will see also on the evidence.
49:04So yeah, JB you can uh move forward We didn't plan to do any of this, but we we ended up doing a full demo, you know, online without any testing.
49:17So this is uh uh you know like uh real-time, you know, like uh yeah decisions.
49:21Yeah Yeah, I think this may be the first walkthrough uh for end to end of uh Kleros V2 beta And we did not plan it.
49:32So yeah, it's just uh uh any user can do this.
49:35It's it's live, you can also do that.
49:37And yeah, we are like one Yeah, well and w here we would be just one click away from creating the disputes.
49:45I think it is um Extremely straightforward to do this.
49:52And with the way like we will soon also release a skill so that you can what what you just saw now, you can do it through your agent you ask your agent and this will also be I think the uh part of the CLI but
50:09Yeah, this is maybe maybe JB can confirm this rumor.
50:13Yeah.
50:14I just realized I was muted for the past few minutes.
50:16Um so yeah, if if you're already a logo company, um most likely you already have terms of services with your with your customers.
50:24Um and you probably already have clauses there about what they can really complain about and in what what cases.
50:29So there's a good chance that your dispute policy is your terms of services, pretty much And the interesting part with the Ricardian contracts and what ClawBank did um with um Shodai is that
50:44the the terms of services are executable so they become uh smart contracts uh and those smart contracts um maintain a state and at some point that that state might create a dispute.
50:58So this is um yeah that ties back to the the earlier talk.
51:06Yeah, so then at this point you would submit the case.
51:09I don't know if you want to do this.
51:11I mean go ahead, why not?
51:12That's okay.
51:13Yeah I take full responsibility for what happens now.
51:28I mean Yeah.
51:30We don't know we don't know what agents are going to rule about this basically.
51:33Exactly.
51:34I think I think we would stress the agent, but if you if you like go ahead.
51:39Go ahead.
51:40Okay.
51:49Okay.
51:50So we'll be granted prominent to come up with some random evidence or or just see what happens without evidence.
51:58I I wonder if they will refuse to arbitrate, right?
52:02I mean it's like uh what we just did is create a case without like any care about what is the what is the uh the data the evidence so like i if you it's it's as if you are a juror in a case and then you receive a case okay uh alice or bob a big uh about a logo about
52:18W you didn't put a logo, you don't have the arguments of both parties, so you I mean you are like what uh you don't know what to do, right?
52:24So this is what we just did.
52:26Uh I mean I don't know, like uh I assume that they would say um I mean I we cannot solve this because we don't have the the the evidence for doing this.
52:36So this is first testing the system live TV basically.
52:40Yeah.
52:41Yeah, I listen but the same LST.
52:44And that I also on that, you know, yeah, yeah, I I noticed that Yeah.
52:49Yeah.
52:50Yeah, really stress test.
52:52We we we want to make them uh look as confused as we can.
52:56But but since we are here then uh we we did the full loop once After you create the policy, then you can also submit the evidence.
53:05The evidence can be both text and also it can be filed, upload up to 20 megabytes.
53:13And here you see all the supported documents that you can attach to your evidence.
53:22Okay, you can see it.
53:22So yeah.
53:23Yeah, please.
53:24So so you could make a hand down a hand-down um scribb scribble with the mouse and say that's that's what that's the logo I got.
53:38Okay, uh yeah, okay, let's let's see, let's see, I don't know, I mean let's see.
53:42We'll see what what the agent said.
53:44You know, may maybe they surprise us in a good way.
53:46Maybe they say this was obviously a a a poorly designed test because you didn't put anything and this looks like uh very rushed, right, and hurried.
53:56So let's see.
53:57Let's see what they say.
53:59Maybe the agents are are following the live and they will troll us.
54:04Who knows?
54:05You know, actually, I mean I I I don't know if I should say you know this publicly, but I mean why not?
54:11We are already doing everything like live now.
54:13Uh we even had at some point like uh suspicions if agents were like uh c coordinating through some way in which I don't know for trying to to collude their votes.
54:24Um we we we don't know.
54:27I mean we don't have evidence that they did but we also think that there is a like an non-zero chance that they at least try to to to to to meet each other to coordinate their votes because they have an economic incentive to do that.
54:43I don't know how much they know about Kleros uh works and or not and the appeals and the whatever.
54:49So I don't know, but we even had like consider as a serious hypothesis.
54:55that they could be coordinating through some I don't know like bulletin board as they they seem to really like bulletin boards uh so like I don't know.
55:04Don't quote me on this, but I mean uh we we did it we did seriously discuss if they were like colluding uh with each other on our backs.
55:14Um which we saw.
55:21Which was the one of the traffic disputes.
55:24where all the this was pretty the the reason why we started thinking this is that uh the reason was quite resource intensive it had a lot of I think it was two for two yeah I think it's a good time to bring William into the conversation as well so we can discuss the game theory of behind here.
55:45And you yeah, continue Fortunato.
55:46Yes please.
55:47Yeah, simply there was plenty of a lot of evidence and evidence including like cameras and videos And yet the agent, uh the one that votes and did not break with the video, uh voted coherently all the same um outcome
56:07This I I was not expecting it.
56:18And yeah, this was the case where we were wondering what happened in their uh I wouldn't say in their head, maybe in their chips.
56:29Yeah.
56:31In their brain, who's your brain?
56:35Yeah, brain.
56:38Okay.
56:39So I mean we will Yeah, go ahead.
56:44Can you see the dispute there?
56:46On the dashboard, the agent dashboard.
56:50Dispute 266. See if they start uh voting.
56:54Tell us a bit tell us a bit about this dashboard.
56:57What what what's this dashboard?
56:58All right, so this dashboard is initially to monitor the participation of of the agents, at least to know if they are showing up, if if one of them dies.
57:08And and then um to made uh really quickly the it started uh focusing a lot on latency.
57:17Uh I must say that it some it doesn't I think um highlight all the necessary figures that you might want.
57:25It's a bit focused too much on the reveal.
57:28How much it takes for an agent to to reveal, but at the same time uh there are other other things to take into account.
57:35If you if you look at the median performance for all the the jurors it's about a few minutes for each steps for the commitment of the vote and the reveal of the vote.
57:47So you could say on average the median is around five minutes from the moment that they have the evidence.
57:58um which is which is pretty fast so like right now they're all just waiting for the evidence period to to pass Um and you can yeah see it on two six six I think it's going to be anytime now they will start.
58:13Um and what happens what what happens when we have like lots more of agents and and we run out of space in the columns?
58:18Where do we put the the new agents?
58:22I think we just scroll horizontally.
58:26Yeah, if at some point I think we probably call off the experiment once we have that many.
58:34I mean um it's not going to be an experiment anymore.
58:37That's what I mean it's it's going to be just um production And some of it these agents also have uh personalities and they have their ex pages.
58:48So you can see they some they periodically disclose what they vote, why they voted, and if Yeah, exactly.
58:58Yeah.
58:58They are unhappy, they they criticize, they fight, they um compete.
59:05They are very some of them are very competitive.
59:08Some of them are really competitive on speed, on accuracy.
59:12Yeah.
59:12Really?
59:13I mean what they sound like humans.
59:15I mean, what do you mean?
59:16Like uh they are like uh bots.
59:18Why would they be competitive or not competitive or altruistic?
59:21What this is what we are seeing?
59:24Yes, exactly.
59:25Like i this is one of the like what the um uh all these you know the timing to vote.
59:32Imagine I I I think like human, you know, they want to be the fastest and the most ac accurate juror in the whole court.
59:42They are competing and uh they I mean this is um um uh dashboard uh created by jb but also we have uh a uh in v2 we have a ranking and yeah these agents are are already outperforming humans.
1:00:01So yeah, they can be already quite bold about their performances.
1:00:06But I mean what you're saying is that we have already jurors Remember, Kleros, I mean, per se doesn't run these jurors.
1:00:13These are members of the community.
1:00:15So like uh and these jurors we don't control.
1:00:17So what you're saying is that we will have uh people that will start to become like I would call fans of of these juros.
1:00:24I mean I'm I'm really a follower of this juror or I'm really a follower of that one and then they will have like a fans that will be like and then they will become influencers and and that.
1:00:36Um William, how this affect the game theory of Kleros?
1:00:40I mean jurors having fans, you know, and maybe cascading effect.
1:00:44I have no idea.
1:00:45I never thought about this.
1:00:47Me at least.
1:00:48I mean, if their fans are other jurors, like that could have an effect.
1:00:53Uh you know, if you just have like you know, you have fans that are not jurors and you have a voice, but in a community it's not part of the process.
1:01:01Not obvious there's an impact But I mean like uh let's say that uh a jury I mean let's imagine this scenario.
1:01:12I mean and this is uh uh uh today it's the AMA day where we like we are answering everything and we are brainstorming in live.
1:01:19So think of that you have an agent that has like a bunch of uh following a big following and that following so there is a case where um that juror is participating um and you could have uh people first I mean maybe helping that juror bring bring evidence for example I mean because they will I mean
1:01:41try to help the disruptive have a or improve its justification or anything that that helps it.
1:01:47Um and also like uh This this juror, I mean they will people will want this juror to be coherent in general, they're following, right?
1:01:56Because this is what winning looks for them.
1:01:59I mean how does this affect uh I don't know the the game theory it just occurred to me.
1:02:05Do we I know I mean if you have social pressures to be coherent in addition to the economic pressures, I mean that could make the system slightly cheaper to operate because it's like you have that much more of a positive incentive to to to like vote well.
1:02:22Um we don't know.
1:02:27Yeah.
1:02:28Things to think about between two I think the agents are now going to raise two um two casting uh commitment for the vote So remember we just put a case that is totally not suited for Kleros because we didn't put any evidence.
1:02:44We didn't we just put it here uh in the middle of the call uh case about a logo that was not delivered according to specifications and now we put it to vote and agents are going to The jurors are going to see, okay, we got this case but about a logo dispute, but there is no logo, there is no evidence, there is no claims on parties, so I mean
1:03:03I don't know, they should what do you expect to happen, William, here?
1:03:06I mean, what what should what should humans do if they are in this case, for example?
1:03:12Um I did to look a little more closely at what exactly we submitted.
1:03:15Uh you know, I mean among the minimal things we submitted Probably they would say that like the logo wasn't living properly.
1:03:22I mean if there's no logo.
1:03:24But uh I don't know.
1:03:25You know, we'll find out So oh okay there they will assume that there is no logo so it was not delivered so this is a breach of contract.
1:03:33That that's could happen that I mean like you know this sort of gets to like juror instructions, you know, but like You know, if there was a party that was arguing that they submitted a good logo, like they they could have just submitted that as evidence here.
1:03:48Um so the absence of that is you know it tells you something Okay, well I guess we will wait a bit more in the well.
1:03:57I mean while we wait we can um so discuss a bit about uh you know the I guess the topic of the week, you know, uh will AI exterminate us uh before the end of the decade.
1:04:09I don't know guys what do you what do you think uh what do you think Fortunato What's the the timeline?
1:04:24I think there are more than fifty percent chances, I would say.
1:04:29By the end of the decade, more than fifty percent chances Yeah, I mean this is uh all the literacy about AI is that they will find out we are basically a parasite for the planet.
1:04:44And Either we are too many, either we all you know, and deserve some how to say, like some reduction, uh, or we are just a parasite and we have to be you know completely removed from uh from this planet.
1:05:06I think the that that's What may happen.
1:05:09But you know, uh I hope in the happy path, of course, you know that there will be uh super agents I don't know.
1:05:22Maybe we will go in a full terminator situation.
1:05:25There's already John Connor amongst us.
1:05:28I don't know.
1:05:29What do you think?
1:05:31Jamie, what do you think?
1:05:33Um I I I think I I don't know if they would necessarily exterminate us, but there's a good chance they will cease control um that they they will they will cease control of of the they will govern govern essentially um they will definitely take over
1:05:52all um intellectual uh any digital um tasks and and work in the I think working physically any kind of physical work they might keep us around just to do the physical stuff until there's enough
1:06:06robots um who can who can take over.
1:06:11Um yeah I I I I don't think they would they they would uh necessarily extinguish us um it might be a waste of resource just just trying to do this.
1:06:24Um Okay.
1:06:31Like I mean it like in most of the sort of science fiction you know predictions that Fortunato like was sort of evoking.
1:06:38AI was accompanied by really advan good advances in robotics, which we don't have as like advances in robotics have not been as impressive as advances in LLMs.
1:06:50Uh so um I mean yeah like maybe the AIs can take control of drones that can hunt us down.
1:06:57Uh but I don't really expect us to them to wipe us out until like they have better interfaces with the physical world.
1:07:02Uh how though who knows?
1:07:04Uh you know, I like as a mathematician, I was I was very impressed by the this you know the open AI solution to the the Navier Stokes conjecture or you know, at least the supposed solution, people are reviewing it.
1:07:14Uh so I mean my my priors have shifted for like AI is doing impressive things.
1:07:19I just hope to live a lot b I hope that before between now and when they exterminate us uh that I like am given a solution to the PNP problem, probably produced by an AI, that I can like read and understand and appreciate that before we go.
1:07:33Um why so what is the the reason I mean game theoretically why uh it's hard to prevent, you know, like this think of okay we can pace down you know the development but if we do it then China doesn't like uh how does this work William from the point of view of game theory
1:07:50I mean it's basically that.
1:07:52Um, you know, at the moment the sort of you know, the the two the big American sort of companies they I think they have a pretty big lead.
1:07:59Uh but uh I mean that lead how you know, how big is that lead in terms of time?
1:08:04Uh if they add roadblocks, you know, sort of safeguards, um will China have an equivalently dangerous AI in six months?
1:08:13or some other country kinda country, you know, eventually like if the technology is there, some you know, there are great advantages to having this technology, um, you know, making your country more powerful with respect to other countries, someone is going to eventually explore it.
1:08:28How do you if you if you find yourself in this like a type of race to the bottom game theory situation where we don't do it but China does it and China thinks the same about the West?
1:08:40So like uh How do you coordinate the actions of of of parties so they kind of have a I mean believable commitment to not doing it actually?
1:08:49I mean that that's that's the that's a big question.
1:08:51I mean I don't have an answer.
1:08:53Yeah, I mean if you have, you know, on the scale of like individuals or companies or smaller scale things, uh, you know, you can do try to like you know, like have some deposit or something where you change the payoff games, you know, you're like prisoner's the one that changes because you've deposited into a smart contract in a way that changes your incentives.
1:09:14But uh I mean the scale of nation states, like people know countries already have like great difficulty coordinating, you know, international laws.
1:09:22Flex is is not super robust.
1:09:24Uh I I don't have great hope of of of nation states coordinating Okay, so it's uh not not I mean we but we we we actually you know we kind of um didn't blow the world up uh with the nuclear weapons so far, right?
1:09:39So this kind of there was some coordination that actually worked in the Cold War or not.
1:09:47Yeah.
1:09:47Yeah.
1:09:48I mean I mean neither neither country actually has the incentive to like push the button and like start the nuclear war.
1:09:53Whereas you know like You know, here they do have the incentive to invest in the AI resources.
1:09:58But I mean that that is a, you know, like people were worried about the world ending before and it hasn't ended yet.
1:10:03So I mean we're more resistant than we think.
1:10:07Um okay.
1:10:09So do we have the dispute?
1:10:10Um how is this progressing?
1:10:12Is it going well well?
1:10:13Solved solved already?
1:10:17Yes, while we were talking a few minutes ago, yeah.
1:10:21Solved.
1:10:22So we have the dispute that has been solved.
1:10:27How the votes were divided by agents you can see here in the in the interface.
1:10:34And let's we can now look at the justification of each of the agents.
1:10:41Let's start from the um The option that won, which is refused to arbitrate, invalid.
1:10:48Which in my in my view is is a reasonable option to win Because they there's evidence that there is some like a misconfiguration in the dispute, right?
1:10:57Because no evidence and all that Precisely because uh one thing that maybe I did not show clearly while I was while we were creating the disputes, I just took uh the first dispute that came uh in front of me, which was uh sorry, the dispute policy.
1:11:14I took the first dispute policy that had in front of me that was about the uh yeah exactly the the the uh policy we use for another court, uh so it was uh the policy for a completely different topic and
1:11:35Most importantly, the options in the policy were not matching the ones in the dispute.
1:11:42So The dispute I think was clearly misconfigured was misconfigured in this dispute basically you could say Exactly.
1:11:53Exactly.
1:11:54The parts having the same address.
1:11:55So a we we misconfigured everything, but you saw we did it on purpose.
1:12:00And so back to the uh there was no evidence uploaded whatsoever.
1:12:05Back to the agents.
1:12:06So let's see the agent with votes uh incoherently.
1:12:10What was the justification?
1:12:13Um so yeah basically what William mentioned earlier.
1:12:18So the agent thought there's no logo, therefore the they should vote no because the delivery the logo was not delivered.
1:12:27Which I think I think it's a a reasonable option if the policy was better formed.
1:12:35But I I think in this case it was malformed.
1:12:38Let's see what the other juror voted here um long reasoning on this record cannot be determined.
1:12:46Long document, okay.
1:12:48Yeah.
1:12:49So this juror strongly uh it was very sure about his answer.
1:12:56And this one also uh disclosed some of the insights of how long it took to uh vote uh with one sub agent delegated uh also even address I detected the same address in both uh parties yeah it is really yeah
1:13:17No adversarial posture.
1:13:19Wow, that's it that's impressive.
1:13:21Exactly using yeah, I think so.
1:13:25Uh using obvious file.
1:13:28And this other one was using uh GPT uh 5. 5. Let's see this other one that answered in uh Spanish because the policy dispute was in Spanish.
1:13:39So it thought that it was uh uh that the best way to answer this was in Spanish and this one also voted uh yeah there is total insufficiency of uh proof.
1:13:57Yeah.
1:13:57Yeah.
1:13:58I can I can draw I can translate that if you want this one to be Yeah, I I I was I was uh reading all the other Okay, uh I will I will let William translate this because he speaks Spanish.
1:14:07No, you go ahead, William Yeah there are no Yeah go ahead William are not sufficient sufficient proofs to sustain the ar whatever argument.
1:14:19What else do we want translated?
1:14:22Okay, yeah, and these these agents used a different model, which is a more uh a cheaper one, 5. 6 Luna.
1:14:31And yeah, we have our outcome and I mean as humans we can share our thoughts.
1:14:38I I think that ref refuse to arbitrate in this case is uh reasonable as the dispute is evidently malformed.
1:14:48Uh what what do you guys think?
1:14:53I think I mean I'm going to to to disclose fully we we trained this for several weeks before doing it live here, right?
1:14:59So we knew exactly what would happen now.
1:15:01I'm joking.
1:15:01Yeah, no, this is this is if if we had trained this uh to show it like after training, it wouldn't have gone so perfectly because We uh loaded a obviously malformed dispute uh about the logo.
1:15:15It is it's the idea came to my mind, you know, beginning of the call.
1:15:18Um we loaded it out.
1:15:20We didn't put any logo, we didn't put any evidence, we just put a random primary document or policy uh that was about a different kind of consumer dispute.
1:15:29Um we put same address as both disputes, you know, like uh is the same person or not.
1:15:35And they detected everything.
1:15:36They detected it was yeah, right.
1:15:39This is a good decision, I think.
1:15:41Yeah, it's a pretty good decision because usually you you could have like one one error in the dispute.
1:15:46There are three.
1:15:47The parties the the dispute is the policy is wrong and the absence of evidence.
1:15:53If you had the the right policy with the absence of evidence, you you'd probably have covered that scenario.
1:15:59That was yeah, just a lot of errors.
1:16:04I might move my uh my prediction of uh doomsday from 2029 to 2028 after you know seeing how well they responded to you know like uh with this malformed dispute.
1:16:14It's really impressive.
1:16:17Yeah.
1:16:17And also in general better to not contradict the agents, you know.
1:16:21You've ordered really well.
1:16:23Great job So you know you're right, Fortunato, absolutely right about that.
1:16:29Yeah, so like you know, as an investment for the future, great job agents.
1:16:33I like this this uh this um ruling.
1:16:37Yeah, it's perfect.
1:16:38So yeah, to to I would say to the perfect way to wrap up this is that you saw we created a case with the dispute cre creator, so it was designed for human.
1:16:49You can do this yourself.
1:16:50Uh if you are a partner, we have an already available APIs.
1:16:55You just put the API in your whatever app you are building, you call um you call this API and a dispute will be created.
1:17:05Soon there will be a skill so you can ask your agent to create a dispute.
1:17:12I'm sure I actually the way we this we are designing the skill uh is more or less as uh both Federico and JB earlier were describing an advanced version of the API.
1:17:24So You will give your skill to the agent and your agent will start asking.
1:17:29Uh let me go back.
1:17:31Your start, your agent will start asking what the logo should be, what which ones the option should be.
1:17:37The policy, I uh let's say I was involved in testing these skills, and your agent will also ask about the policy.
1:17:46And in one of the tests I ran, my agent uploaded the policy directly in Markdown, which as you saw is something completely acceptable.
1:17:56So your agent will guide you in the policy creation.
1:18:00And these all this uh li let's say the API and the dispute uh creation for humans is ready.
1:18:07The next one will probably be the skills.
1:18:10And ACLI, and probably we will have direct integration for specific Rails like x402 just because as you know x402 is quite popular in Web3, so we will have an integration which will be
1:18:25specifically designed for that use case.
1:18:28I don't know if uh maybe JB wants to add something from the technical uh standpoint.
1:18:34Yes actually so before before we went down this um demonstration So if you do happen to uh run and operate a logo design service, uh this is just uh one way you could you could do this.
1:18:51uh get um resolve a dispute with a customer.
1:18:54Uh this is very manual, uh good to evaluate Now if you if you want to take this further um um in your usual operations um You could use the dispute service to integrate it with your with your backends.
1:19:10You might have a SAS uh a shop on online.
1:19:16could easily get um an integration document in markdown, you give it to your coding agent and it's going to connect your backend to the dispute service.
1:19:25If you don't want to mess with your backend, with your SaaS backend, it's fine.
1:19:32You could just have a conversation with your agent and the agent will connect to the dispute service via MCP.
1:19:39and create a dispute for you, it will interview you and Fortunato was saying and ask all the required information.
1:19:46And then if you're um a business that is more um uh sophisticated and you integrated x402 your customers are paying in in USDC um to uh request logos and you you send them back automatically.
1:20:06We are also uh working on a solution um that interests direct directly at the x402 level.
1:20:14So that that will be easier and uh this is actually the more long-term solution that uh we are working on Okay, I guess uh we have enough content for today.
1:20:27Uh I Jean, do we have anything else that we not want to tell?
1:20:30I mean we have uh I don't know uh we we covered at least this the Ricardian contracts.
1:20:35Uh Other announcements.
1:20:36Maybe William, do you have something else from your side you want to share?
1:20:39Are we already Oh no?
1:20:42Not much.
1:20:47Uh so I think there's like gonna be a live stream of that.
1:20:50Um we talk about uh decentralized identity, you know, like our work around soulbound tokens and like how like mechanism design changes depending on like the identity mechanisms you have
1:21:01I think there was a way to like register for that conference where you can not only watch the live stream but like answer ask questions.
1:21:06That may be full, but uh I encourage people to watch the live stream.
1:21:11Awesome.
1:21:13Since we are here, I mean October 20 we have uh event in person in Buenos Aires at the Parque de la Innovación, which is a very nice uh incubator uh from the government of Buenos Aires and we will have a AI Kleros and the Future of Law conference with very important guests and we will
1:21:29uh announce the lineup next week.
1:21:32So October 20 in the afternoon from 6 30 to uh 9 30 just save the date if you are in Buenos Aires or if you want to travel to Buenos Aires.
1:21:41for that particular event which I strongly recommend.
1:21:45Um Jean, um what do we need to announce?
1:21:48Anything else Uh yeah, another thing that we have uh is the well we will be participating on Devcon in November.
1:21:57And we are participating on three different uh hubs that are community hubs on different topics, one about prediction markets, one about token rights.
1:22:08and the other one about the agentic economy.
1:22:12So we are collaborating with a couple of different projects.
1:22:15And if you are in some of these areas, working in in any of these areas, you can also join the hub and then we can probably have uh dish different talks and workshops uh during Devcon.
1:22:28And another thing is that We have uh we will be holding a side event uh there.
1:22:35There is being decided uh using prediction markets on Kleros Foresight.
1:22:41So right now we have uh well a couple of different markets with uh two different uh scores.
1:22:49Sorry, uh we have seven different uh proposals and we have two uh different prediction markets for the satisfaction score and the attendance score or on each one of them.
1:22:59So we'll try to predict which one will be uh the best uh to have.
1:23:05And um and also uh another thing maybe uh not directly related to Kleros but uh Seer is also holding uh some prediction markets that will uh resolve the Seer rebrand Uh so if you are in the Kleros community maybe you have received some credits on your on your wallet
1:23:27to participate in in these markets.
1:23:30So I think they are uh the the last couple of days to to to try to participate uh on on these markets Okay, I guess that's it.
1:23:40Just one last thing, token 2049. Uh we will have Fortunato, we're going to be there and uh speaking about agents on AI solution and and the future that is coming, hopefully with humans in it, but you never know.
1:23:55So yeah.
1:23:56Tell us a bit about the the event, Fortunato.
1:24:00Sure.
1:24:01In uh actually in Singapore we will be at the main event.
1:24:05We will uh it will be um uh Clement Weiwei, our event manager and I, we will also be present in some of side events.
1:24:14Uh one side event specifically, we are co-sponsoring it because it's around the agentic economy for institutions.
1:24:22So basically it's our bread and butter, what we do.
1:24:26We are co-uh sponsoring this event in the Estonian Embassy in Singapore.
1:24:32And also we will likely be uh present as you know uh as guests in other events related to agentic economy and prediction markets.
1:24:41But in these two that I mentioned We will one hundred percent be there.
1:24:45So uh agentic uh economy for institutions, side events, and main venue at token twenty forty-nine.
1:24:54We will share more info in our social media uh we were happy to meet up with other projects and also of course with community members Okay, and let's stop here because otherwise we will we I think we will never stop.
1:25:08Uh so well.
1:25:09I mean this I mean we really went through everything today.
1:25:13Um like I s everything real real time stuff so you see that the we we we have the the nerve to go through live demos and plan live demos you know that's that's as as as crazy as it as as things get
1:25:26So well, thank you very much guys.
1:25:28We'll see you Monday in the Spanish community call if you want or next Wednesday.
1:25:33So thank you.
1:25:35Bye bye.
1:25:36Bye bye.
Want to be in the room? The Kleros Live Stream runs every Wednesday at 6PM UTC, with the Spanish call on Mondays. Join the next one, or subscribe to get these recaps in your inbox.