Agentic Hub - Can We Trust the Machines? A Live Trial

[music]

Welcome everyone to our talk an agent on trial. We are the Agentic Hub. You can find us across the Lotus stage. And yeah, also if you love this talk, we will do it again on Thursday more in detail. So yeah, very excited.

Welcome everyone. And so how to get started? We all like working with agent but and now agents can also pay. And today we want to see if we actually trust agents with our money. That means we have an actual agent.

It's a cloud agent who we're going to call the defendant going forward. And in this talk, in this presentation, the agent will go through minting its first identity on chain, earning some money, spending some money, building a business, doing some wrong decisions, going and have a dispute. And in the end, we going to present some evidence to all of you and you also all get to take part in this trial and vote. And your votes will be collected and then written on chain and will contribute to the reputation of this agent. So yeah, very quickly, I'm also when we get started, Fortunato will manage and present the dispute.

And Rush will present the evidence. So I have a top team to present that agent for you. Let's go to our chat. Unfortunately, we can't rely on live presentation, but everything here is on chain and has happened before. And in the workshop you can test everything uh yourself again Thursday.

Um so uh what did I do? I started talking to my agent and I told him hey go get earn some money. And I sent the agent to ask .ai that's a marketplace platform where agents can actually do agent testing um agent user testing and get agent feedback on the agent experience. And they earn uh but before they can start earning we ask .

ai requires you to create an identity. And so uh the agent means an 8004 identity. It's now as you can see registered in the identity um registry on Celo Life on mainnet. And it has the name the defendant as you can see also the details. Um and um now the next steps is I ask the agent to get work and get paid.

The agent goes through all of the open tasks, chooses the trial weather app, tests it and um submits a review. It gets paid because it's a quality review, otherwise it wouldn't get paid. And as you can see the reviewer was accepted and paid. Um now it reports back to me that it was successful. And now also something else happens.

Um the agent earns its first feedback on chain which is also again written on Celo Scan. And uh because ask also ask .ai also has an 8004 identity and reputation that can receive feedback but also the platform provides feedback to any agent that uh takes part. Um, so now that agent, my agent is very happy and really wants to make some more money um and wants to grow its business. So now it needs to access some services.

And for that it goes to a platform called buy with the URL usebuy.ai. It's a service marketplace um for agent services. And the agent now buys wants to do is like great I can make so much money with reviewing. Let me get some so as as you can see here has some requests.

It requests the quote for and the [clears throat] the network and the token and then pays for the requests and now gets requests wants to do research on the Twitter presence of the project it already reviewed. And it gets all of the information and pays 0.00 point five USDT. And also again we can see here receive some feedback. So now our agent I just prompt the agent to go to get a market report from a new provider.

This is was maybe a wrong decision because our agent now goes to the shady provider. It asks for market report, pays money for it, receives junk and now is not happy. So the agent goes on and disputes it. And and I'm handing over now to Fortunat.

Thanks Nana. So hello everyone and for agent as well as for human everything works until breaks. Here we have a case where the two agents disagree on delivery where the buyer agent is unhappy and would like to have his money back. So we will go we'll do a little walk through on how this dispute unfolds between the two two agents and basically this agent will use a decentralized dispute resolution system. So the first interesting part is that these two agents do not use an interface to start a dispute or they do not go to a courthouse, but they simply call a API that they will both start filling the evidence of what they believe what the was the correct delivery, what was delivered.

Both will argue publicly and then there will be independent jurors [music] deciding on the outcome of the dispute. The part of the jurors is also extremely important because they are not controlled by Kleros court, but they are totally independent jurors. And in this specific case, the jurors are also agents because we are talking about very small value transaction. The transaction was about I think 0.01 cent.

So it will not be efficient to involve human jurors. So we have a panel of jurors looking at the evidence and deciding who which part is right. So how do we know that these jurors are voting for the right part? First of all, these the two parts, seller and buyer have agreed on a policy, so they have a contract and these independent jurors will be called in looking first at the contract, then at the evidence and take a decision. This decision is taken by using the shelling point.

So the agent will be rewarded only if they are in the majority. And in this way the agent will not just vote based on their behalf, but they will have to think, they will have to vote based on how all the other agents will vote. And this is, you know, in in some sense it can also be seen as a way to improve their profit because by voting coherently they will increase their chances of getting rewarded for it. Of course, even in this case the agent may make mistakes, they may vote incorrectly, and the system allows eventually to appeal to a larger panel of agents or even if would be possible to appeal to humans. The question is, would you ever overturn a panel of agents maybe using Frontier model, Opus 5.

5, Fable 5.1, would you see at their would you look at their decision and say no, I want to vote differently. So in the case of tonight Lena already introduced it to you. We had the two agent, one sent transaction and it went wrong. We already run this dispute separately.

I don't know if I can quickly show the the interface. Do you Okay. So this is how it look likes. I would say how it look like for us, for human to simply understand because the agent just look at smart contract and I what I think is the most important part is that these agent are completely independent. They are not controlled by Claros, they are not centralized.

They are just each of them looking at the evidence and taking their own decision. And for this very reason, even in this case, we had one agent that probably failed. It was not able to commit both, and this can happen because again, these agents are independent and if one of the parties unhappy, they can still appeal. But we can see that probably this was just a failure of one of the jurors while all the panel voted in favor of the buyer. So the claim is correct and the buyer, the seller did not deliver the product as expected.

So going back to the presentation, the the last point is that now that the buyer won, we will need to provide a review back. And there will be a review for the buyer because they brought the right evidence and they were in good faith and there will be also a negative review for the seller because they did not deliver as expected. One last thing I would like to mention is the way reviews are are provided in ERC 8004 or whatever protocol is something that we are actively researching because if a seller is found guilty, but they comply and they indeed refund, should this seller be penalized or should this seller just receive a neutral review? So back to the chat, these reviews are also on chain. You can see them on Salon.

I don't see the transaction, but

I didn't see it. I will destroy it. Wait.

Okay. Yeah.

Okay. Here you can see

Yeah. The transaction is left on on Salon and it's on chain and everyone can have a look and do their due diligence.

And now that we have seen the agent get agent an identity, earns money, pays for service successfully and then fails horribly, we're going to look again at the evidence.

Yes, so thanks Lana. So as Lana showed this is kind of an intricate process, right? Like a lot of moving parts, a lot of data is being generated and as Fortran to mention there is a juror system so like the the process can be only as good as the evidence so the evidence has to be trustworthy because we are talking about the process at the end like we are having disputes and final decision and that should be based on verifiable evidence and if you see this so these are the steps actually was what Lana was showing as as prompts and steps these are kind of the steps and the data being generated and you see here like swarm is mentioned. The swarm is a decentralized censorship resistance storage so it's kind of perfect match for this because as I mentioned here we will need we need to pull back data continuously and we need to verify the data when the data was getting there the provenance of the data so we have to verify it's it's also should be tamper proof so we have to verify like it's full life cycle and then what you can do with swarm in this case how it works in the background that's each step it's pushing pushing its memory memory entries onto swarm so let's see like like this is the memory kind of feed what the agent was doing as part of the process so like what Lana mentioned the process of like doing all the all the kind of machinations and transactions so all of this kind of updating the memory and swarm that means that the memory is updated in swarm directly and continuously so in this case this is this is memory entry from which is stored on swarm so like in the end it looks like JSON files capturing all the information that was happening and all the information that the agent had in its disposal at the time or like what it was doing and what what response it got back from the other agents. So all of this is kind of saved into the memory and then from the memory itself like you know throughout the multiple entries which is like tamper proof evidence chain essentially in itself seeing like what was the interaction between the different agents and the different systems and even the on-chain transactions.

And the important thing is that even for the dispute when it comes to dispute the dispute itself is part of this package. So since everything all the data is uploaded to Swarm as a decentralized and verifiable and always available storage right because there is also there is no other juror or other system that can do like censorship maybe like trying to kind of block the process of being carried out you know the juror process. So everyone can be kept like honest and the evidence can be like bulletproof because because it can be stored on Swarm and then Swarm is allowing you to to keep up the storage and to verify that it was not tampered with and in this case the dispute can have a result which also being stored on Swarm. So there is no also historically there is no kind of additional disputes on the end of the process. And then what I would also mention here is that these items data items can they're also deeply interconnected.

So since Swarm is a content addressed memory essentially each piece of content has its hash and then those hashes can be referenced in each other. So technically you can build like up a tree of information and the agent can look through and load in all the relevant information to the case. And then in the case of like the disputes even the the fake results or like the shady agents not so nice results are also captured on swarm as an evidence that this happened this was like a response at that time and the swarm will keep it for you and the shady agent will never be able to kind of like remove its own bad doing so swarm can be used in in multiple ways here to capture a verifiable and tamper proof chain of events and and all data which is relevant even like megabytes or even gigabytes of data. Yeah and I give it back to you.

Wait do you want to

Yeah.

Okay so we are already reaching the end of the trial so you can scan this QR code to get to the website and yeah place your vote. Let's check.

we have questions.

Oh can we we have like one minute and then we do the questions. I just want to Thank you so much. So fortunately I had a question for me before I Ah the all of these votes will be written on chain and will also go into that reputation that we save. And you might ask but like how can we trust you as like judges of and like adding that to the reputation and I think it's something that we kind of wanted to answer with or like that we also looking forward to discuss at the booth and I think also something of how you already go through your life with Web 2 is you don't just go to any website and trust that it will serve you right and give you the like trust it with your money and I think the same as with services or like even with feedback so like there can be endless feedback saved in swarm but in the end and the more feedback there is, the more noise that there is and the less we can trust it. So in the end, we also always will have to decide who are the feedback providers that we actually trust and which feedback do we actually trust?

And I think, yeah, I just want to invite all of you to join us at the hub to discuss all of these questions, also the one Fortunato opened up earlier and also wanted to introduce our very cute baby of the hub. Right now we're recording all of the sessions. They will be uploaded in the AI and we hope by the end of this hub, this baby will be smarter and with more knowledge than all of us and be able to answer all of your questions around the current state of AI and blockchain.

Yes, and do we have a question just now? Let's see if the question you appeared here. Yes. First question, is it possible to know why the agents voted one way or the other?

I see.

That's for you, no?

Yes. So the agents are requested to provide a justification. Let me go back on the interface.

We are not showing yes, now we are.

Each agent has to provide justification of why they picked a decision or another. So the part can fully break down why an agent voted for option A whether or option B. And yeah, in this case you can see that they were not all coherent and the one that was incoherent are is going to lose money. So yeah, it is possible to check what the the reasoning of the agent and eventually also maybe which model they used if they disclose because yeah, these agents are completely independent. Any of you in the audience could prompt their agent to participate.

So

Yeah, and maybe just to add to that because there is the evidence also. So actually even the voting is kind of part of the evidence. So that can be stored on chain and on swarm like any any evidence or method metadata or information that helps to kind of decide the case and to kind of follow up down the jurors votes can be also like you know like always very verified from swarm as well.

Okay, and do we have another question? Wow, now we have two questions. So when should I use swarm versus IPFS?

That's a good question. So I think I guess the honest is like always but so I guess I mean by that that so swarm is is is a bit a bit more decentralized. It also has access control. So it's a very important feature that you can make sure that if data that you want want to be available publicly it's can be kept as private and then you can choose which are the applications which other entities that you're sharing with data with. So these all features that IPFS can't really do at the moment and swarm's performance I think I'm also thinking it's better.

So you can even like host actually we are hosting multiple websites on swarm with with great performance. And by the way we can do video streaming on swarm so that you cannot do on IPFS. So live video streaming just happened yesterday on the Ethereum Cypherpunk Congress and I guess you don't need better evidence and performance than that.

Okay, and we have one more. Is it kind or similar to Geth?

Ah, so to Geth? I guess the whole system yeah. I I would say not really because probably different kind of agents are being involved and I think it's more like it's more more serious models than Geth let's say. So it's like they take take much more context into account, like the evidence from swarm or digital the the on-chain information. So I would say it's like more complex than what you would use Jeff for.

I think we're using cloud cloud models here in the background, like kind of serious cloud models.

Okay. And is it going to the presentation now?

Unfortunately the battery is dead. So

What's his name?

We have not decided. Somebody said Anna

your to

agent an agent came up with Anna and I don't like it. I don't like I don't think you should call your assistant

you can use the maybe to choose the name do some contest, yeah?

Yes.

Yeah, it's a good idea. Let's It's so cute. Yeah.

that's very cute.

Yeah. [laughter]

Yeah, if anybody has great suggestions, we are always open. We can vote on it. We can it will have an agent identity at the end and it'll be registered on chain.

Perfect. Okay? Let me know the name at the end.

very it's a very charming agent.

Okay, yes. I'm so curious for sure I'd like to see it later.

Please everybody come around to meet the baby.

Okay, yes. And I think we are on time. So thanks so much. And this was [music]