Claws Out Summit | Drew Mailen - Valiron

Transcript

Hey everybody, welcome back. If you are interested in agent trust infrastructure, please make your way to Prosperia because up next we have Drew Malin, the CEO and founder of Valerin, the onchain trust and reputation infrastructure for all of AI. So if you're interested in agents working with humans in a verifiable and reputable way, come over here and come listen to Drew. Good luck. >> Hey, thank you very much. So yeah, I'm

Drew. I'm the co-founder of Valeron and we protect infrastructure from bad agents. So, just a little bit about me. Uh, I've been working at crypto for about seven or eight years now. I was at OKX AWS on their blockchain team. Uh, I was at ETH Denver. I was an organizer here and I worked on a project called NFL Rivals. And then if there are any other Frontier Tower members in the audience, I also hang out on the 15th floor on the Ethereum floor. So, come say hi if you're there. uh my co-founder uh Vots who's in Montreal right now he's a maintainer of UNO and worked on conversational AI at Microsoft so basically how Valeron was founded was uh I had an idea for Valeron as a Moody style grading system for X42 endpoints and I wanted to do this work a couple

months ago at a YC hackathon uh Y cominator hackathon and so I scoped it out for a front-end engineer and a backend engineer uh for the architecture and bots and then our other friend mall uh helped out and and we came we were finalists for the the agentic payments hackathon in November. So we just continued to do it and do it and and keep pushing. So about two two three months later we're still working on it. Uh and we also did 8004 Santa. So if you were uh if you participated in that we we did that too. So uh right now there's a big gap in the agentic risk uh space.

We have machine speed chaos. So basically by that I mean uh AI agents hit APIs and they cause retry storms, loops, they rack up expensive fees on infrastructure and this infrastructure is built for humans. Uh so we're not necessarily ready for the machine to machine economy. Uh and this forces infrastructure providers into a binary trap where they're either forced to block all agents uh where they lose their revenue or they could accept all agents and that risks systemwide outages. So API providers need a granular system with trust signals that don't uh force binary decisions.

So there's two parts of our infrastru in infrastructure stack. Uh have any of you heard of X42 and 8004? Raise your hand. Okay. Thought so. Uh yeah. So X42 is just a really uh emerging protocol that allows pay-per-click API calls. Instead of a flat rate fee for uh API subscriptions, you could basically put uh an API behind an HTTP header and then once the header is is paid, you can unlock the pay-per-click API with just one shot instead of spending a ton of money on flat rate based uh API fees. And then 8004 uh I know that's really popular this year here. Uh it just came out in Ethereum with the LA within the last couple weeks on mainnet uh but that's a standard for agentic

reputation. uh it has over 35,000 registered agents across a bunch of different chains. Uh so that allows for feedback score. So an agent could authorize uh another party to give it feedback. Uh and the feedback could be anything from timeliness to accuracy to reliability. So different standards for uh understanding the reputation of an agent. Uh then there's an onchain identity verification component and there's the immutable behavioral history component. So together these protocols create the foundation for a secure and scalable agentic commerce system but we need a little bit more trust in order to get down to uh understanding these agents. So here's some stats. I took these today because I woke up this morning not knowing that I would give this talk and so I made sure that all these stats were very much up to date

because this presentation was actually created for our YC application a couple weeks ago and 8004 is so quickly emerging that the stats just go out of date so quickly and same with X42. So basically uh what you're seeing here is uh the feedback count and average score correlation. Uh then to the left it's not sorry I thought it would be a little bit bigger but it's it's basically showing the different chain breakdowns of feedback scores. So bass uh is the top with I can't even see it here on the screen. Base is the top with almost 50% of the feedback uh that's given to 8004 agents followed closely behind by Monad.

Uh 28,000 pieces of feedback in total. uh average score of 89.8. Here's some stats from the last 30 days from X402. Uh 6.66 million transactions, uh 1.53 million of volume, uh 28,000 buyers, uh 6,000 sellers. So, the buyers are really outweighing the sellers at this point. So, you can imagine the havoc that that's that's putting pressure on on the sell side. So, Claw Hub, uh Claw Hub is basically, do you guys know Claw Hub? Yeah, CL we're we're claws out. So, I would safe to say that we know Claw Hub in here. Yeah, Claw Hub is a skill marketplace essentially for agents where uh agents can go and acquire different skills. Right now, uh

there's 9100 different skills in there and there's ways to see and understand the risk that each agent has. Uh basically every skill has a virus total uh which looks like this from like 64 65 different uh threat providers and they uh conglomerate their scores and uh it says whether or not it's like low, medium or high risk uh from the open claw uh confidence booster. So uh couple different elements of behavioral intelligence. So, uh, 92% of, uh, agents have a compliance issue. So, that leaves like 8% of agents that are risky. There's, uh, malicious threats

in, uh, way too many agents. Out of the 9,000 or so, uh, agent skills that are on the Claw Hub marketplace, uh, hundreds, if not a clear shot of like a thousand are actually malicious or uh, could potentially leak credentials. So, it's a big problem right now that like a lot of people know what OpenClaw is, but they don't necessarily want to do it because they're afraid of what it could do for their infrastructure. So, Valeron is essentially a proof of behavior firewall. Uh, so we do uh two things. We we basically have an onchain score and so an agent has an 8004 ID. We check the onchain reputation of the agent ID and then we see how many

feedback how much feedback it has and we give it insert production access if it has above a certain level of feedback. Uh or we send it to a behavioral sandbox if it doesn't have enough onchain reputation. So the unknown agent arrives, it gets route to the sandbox and then we track behavior uh for things like rate limit uh loop rates uh and request volume among a few other things in the proprietary scoring algorithm that we put together. And then it's moody style. So we have like essentially nine different tiers from like AAA down to C with AAA being the highest quality of agent down to C being like totally blocked. Like you don't want that in your infrastructure. So I'm going to do a live demo.

So essentially like I'm just going to put an API endpoint in here. Uh the endpoint is an X42 endpoint. just have this list here to make it easier for me. But these are all production level on Ethereum mainet. So this is like the flow of how uh the requests happen. So the AI agent sends the request. It goes through the Valeron edge proxy or edge proxy. A decisions made uh the X42 payment is made and the API gives a response. So here's just a couple agents that are actual live agents on Ethereum mainet. So uh the edge proxy gets called.

>> Sorry, I'm just going to try this again. Live demos. Got to love it. Swear to cloud if this doesn't work. [snorts] Yep. Let's try it one more time. There we go. All right. So, uh, this is a good agent that was routed to production. And basically what you're

seeing here is an X42 endpoint on orthogonal. Uh Orthogonal is a uh discoverability layer for X42. So they work with a lot of different uh web 2, web 3 teams. They're actually in the current YC batch at Y Combinator. Uh they're almost done with it. Uh but they uh add a discoverability layer to X42 endpoint. So we work with them to either uh to to basically grade uh endpoints. So this endpoint was a good agent. So it gets sent to production. Oh, and I don't think I mentioned that the endpoint that we're checking right now is weather in San Francisco. So basically, you just have to put the longitude and latitude into uh the the X42 endpoint and then it gives you what

the weather is or the agent what the weather is. Let's check another endpoint. Okay, this one was sandboxed. This one also sandboxed. And then you could see this um you could see this on 8004can.io essentially that like this is actually happening here. This is live uh and this is uh the score that it has. It has five elements of feedback. This agent that we just showed 254 64 um was just being used here.

So it had you know score of 50. No offense to that agent, but it didn't make it to production. So that's that's essentially uh how we work. You know, uh you can put any X42 endpoint uh and understand if you should have agents that interact with it or if they should get blocked. Uh this will save enterprises money. This will save infrastructure providers money because otherwise you're going to be dealing with uh retry storms loops that just raise havoc increase your AWS bill and nobody nobody wants that. So as X42 endpoints come online and I really think that X42 is the killer use case in this like agentic payment economy uh because of the the microp payment aspect like

nothing is is fast enough to execute transactions within like an LLM uh than uh stable coins. This is why you're seeing like Stripe uh put out money for Bridge and you're seeing all these stable coins and USDC come online to to participate in this X42 economy because agents need the fastest rails possible because they want speed and they prioritize speed. If a if an agent is if an LLM is calling uh 50 to 200 different tasks per call per per task then it's going to need that level of speed in order to understand the infrastructure understand the risk level.

So that is uh Valeron. We have our SDK live as of a couple days ago. You could find it uh right here actually to the right. Uh it's live on npm and that's our Twitter. Uh I'll be I'll be hanging out around here if you have any questions. But yeah, thank you all so much for for coming to the to the talk today and uh would love to talk to some of you after the panel's over. After the talk is over. This isn't a panel. Claude makes it a panel. So thank you.