Speaker Panel I Infra For Agentic Apps I Agentic Ethereum 2025 Summit

[Music] in the world of AI in infrastructure and just understanding what's happening in the world of being building things on chain uh join us so I'm going to bring on dear sha Tov DC Andrew Jasper Mark and Chase to uh share about their thoughts on how to be build better infastructure for gentic caps so it'll be an amazing panel discussion and uh we'll get everybody set up now [Music] all right massive world is great all right this is people coming in and uh D Mars I'm gonna hand this off to you so uh he'll be the moderator and uh looking forward to this awesome chat so feel fre to get started sweet yeah super excited we've got quite a stacked panel um so yeah welcome to the eth global panel on infra for gentic apps um to set the stage a little bit um in the last few weeks and months it's become consensus that AI agents aren't just sci-fi they're creeping into our everyday reality making decisions some executing transactions and interacting with decentralized infrastructure that we're only beginning to understand um and the question isn't just what can AI agents do today but how can we Design Systems that truly enable them and before we dive in let's meet our panelists I'll have each each of you introduce yourself and if you want to share one or two sentences on how your work relates to agenta gaps that'd be great I'll kick it off with uh let's start with DC awesome uh so I'm DC Builder I'm a research engineer at the world Foundation um so world is a project that sort of Started from the AI side of things um where it was co-funded by samman Alex Blan and a few other people um so essentially the the the the the way that it relates on to this this problem space that we're discussing now is that we're building the text stack that allows you to prove that you're unique human on chain and online which is sort of like a necessary tool to sort of um enable the internet to still be usable even after the dead internet Theory sort of comes along where AIS just like are able to do anything that humans can do but humans should be able to use everything just just as just as well and there's a lot of also like a lot of interesting things around economics there uh and just general identity that are really interesting and you still want humans to have the self sovereignty rights to access finance and identity in a private way so that's how relates cool let's kick it off to Mark then after that hey my name is Mark and I'm a co-founder at op Labs you know we build block space we're working on the op stack meant to you know commoditize rollups so yeah the the idea would be to you know create abundant block space for these agents to to use and consume and uh you know hopefully people build some fun mechanisms on top and we can see how the agents all you know work together fantastic David hi um I'm David um my background is um I was doing cor research um on ethereum um mainly working on fee Market me um and then a little more than a year ago we started doing research on like uh using ethereum as a commitment device coordination device for AIS uh we were exploring this concept defining this concept of AI M um yeah and then about a year ago um I also started an AI project called P so we are building a personal AI agent so like uh this is like the real AI side where we build like uh agents that have like um full contest on like um personal data for uh people and then we use the blockchain for what I think it's uh very um uh useful for which is trust and uh coordination so I'm happy to dive uh deeper on these uh during the panel and um yeah so we are integrating with a few of the projects that are already on it eum L2 uh like virtuals uh World coin uh for um identification so I feel that uh the ecosystem is already growing and we should build on one another make it even stronger cool and then we can just do Shaw Chase and Jasper in that order sure um my name is Shaw I made Eliza which is an open source agent framework um many people use it to make slot bots on Twitter but I do think it has a lot more capability than that um and currently working on the next version of that um so expect that pretty soon everyone I'm Chase I lead Ai and infrastructure at offchain Labs uh we're the Builders of the arbitrum tech stack so that's arbitrum orbit L1 l2s and l3s and um the arbitrum one main chain that many of you know is the Hub of defi my goal uh in growing the a ecosystem on arbitrum is largely to lower the barrier to entry and improve crypto ux particularly around Defi and gaming oh we cannot hear you Jasper can you hear me now yes we can wow nice save that's quick turnaround hey everyone uh Jasper I'm the CEO and Co founder of hyperbolic I did my math PhD at UC Berkeley I finish in two years which make me the fastest person to finish a PhD in the history of Berkeley I also won a few go medals in various mathematical competitions and before finding H Labs I work at Avalanche and C Securities so uh hyperbolic is the decentralized AI Cloud for ai's open Future so we work with open source Community we provide access to compute and we try to democratize uh AI innovation and accelerate AI Evolution so in the future if we believe that AI agents will will be able to do most works that human can do then uh the the main users of cloud won't be just human it will it will mainly be agents and we're designing the infrastructure to make sure that agents have access to compute and can do all these different crazy uh capability stuff like using tools even like self Evolution and self replication amazing uh well we have quite a panel of experts uh I'm excited for this um let's just Dive Right In I think we can skip a lot of the definitions I think people can see some of the previous talks if they're interested in those but the first question I want to pose to everyone here is what are blockchains and in particular ethereum useful for when designing agentic applications and what are they not useful for as well is maybe the counter question I think there's probably a like a whole Space of different products we can build but I think the most obvious one that I've heard from a lot of people is something like a registry um and that registry could then quickly become like a Marketplace and I think that serves kind of um the value like instead of us in you know we each have our own agents sure but probably have some some kind of app on our phone like chbt or like Google and apple finally get their together and their assistants are actually like useful and they could like go and contact UM other services like instead of having to have everything just right there in the app they could say like okay well let's go actually discover services that we want to use to extend the functionality like can we actually just go use another API without having to write the API interupt to every single API that exists and let people kind of just like go and say well hey I've got it I've got an app that solves that um I think people have pitch this as like five or for agents as one version of this but usually to me it starts with like well we need to be able to find the agents we probably want to build trust and Merit scores around them in some way like hey is this actually useful does it do what I what I want it to do and then to me the natural progression there is towards some kind of marketplace like can I hire this agent to do a thing for me um and and really it's to me it's all just like wrappers around uh the combination of apis and llms but it is like solving a really annoying problem that we've had with apis generally which is they don't write the interrupts between each other and that becomes quickly like a you know kind of quadratic explosion of of interoperability uh that's that's not tenable um and so by having this like reduced to a fair like a form like okay we have the ability to negotiate this in roughly natural language and send information over that's that's been um uh you know like Whit list to send or whatever like your public key or Paving information or like some files um then you can have this like way of um like discovering and extending the capability of an agent without having to write it all and so generally like if if someone's like telling me about their multi-agent system they built I'd be like okay well it's just you building all of them so it's probably just one agent but in a world where we want to be able to like Leverage everybody's uh apps kind of all at once I could imagine this sort of registry thing being really crucial so there's probably a whole lot of other ways and people are building agents into wallets and doing this kind of you know user side and Etc but the the thing that I see is like what is the L1 of agents that that really connects it all together probably starts with that registry yeah I want to kind of like uh agree with this um I feel that uh on chain registry like actually give you much more uh so for I can give an examp example like P we're using Registries for two things one is like we have this data connector protocol which is like based on a t Network people like can uh give o like to the te and like start like fetching their own data uh from web two and uh for this we need the registry because uh we also need like uh Economic Security for uh who runs this data connector um and then the second one is like registry of actual agentic application so the design we have is that like there is a small llm on user device but this cannot do everything it's going to like have some capability mainly around privacy preserving and some API calling and then is calling like uh uh agent Services uh in the cloud right but these agent services are registered on chain and uh what you can do you can do Discovery as show was talking about you can also do uh um uh security right and uh maybe like for the low stakes agent you just do some reputation based on like the visibility you have from the past actions on chain like uh but maybe like for uh higher Stakes agent that like handle more user funds you want some crypto Economic Security you can program that in the registry and then uh uh for like very high stakes if you imagine like in the future there is like uh some onchain MD that does a diagnosis then you want full verifiability maybe yeah so yeah this is exciting and like um uh we can integrate with like existing registry already this is what uh we are trying to design uh yeah with virals yeah I totally agree I think like so what sh and deit dat mentioned about is about the registry but besides that I think like trust and payment is also very helpful to be hosted on ethereum so for example like right now all these AI agents that just like like run on top of like a centralized API but how do you make sure that they're not trying to rock humans or like it's not controlled by any human as as as well and so uh ethereum is the decentralized TR layer and so there there naturally you can utilize etherum to build CH for the a agents right like for hyperbolic we invented like proof of sampling which is a sampling based verification mechanism for a models and a agents and so uh we we we need to use like uh stake and on ethereum to to kind of secure all these AI models and a agents but there there are other methods uh for example like op you can also utilize ethereum as the the like slashing uh as the place to do the slashing stuff um and besides that I mean naturally all these agents will need to pay for their their own inference and do transaction with each other in the future and so uh ethereum can be served as a payment layer for agents largely agree with everything everybody else said on the panel um the thing that makes ethereum particularly special for agents is the ability to create customized and opinionated block space through layer twos and layer 3es so this registry of agents that everybody has talked about so far could live on its own chain that's far far cheaper than using the ethereum main chain nor does IT Crowd the ethereum main chain these agents transacting with one another negotiating with one another talking to one another etc etc um we have some we have some arbiture morbit chains that are doing this right now like capex and um Skynet for AI uh really interesting lots of you know lot lots of customizations that can be made but they can still interoperate with you know L1 with the ethereum L1 obviously with arbitrum with optimism with base um you know they can control private keys from from any chain but they this registry can live on its own you know L2 or L3 okay I see um I I think that's a that's actually an interesting point to double click on so there's a couple things I heard here is like one is Discovery two is Trust Three is payment four is interrup some of those things aren't mutually exclusive and they sort of feed into each other but maybe we could double click on what what does make a a chain good for discovery for these AI agents is it you know is it some L2 that like Mega scales or is it like an L1 that is like maximally decentralized I think there's many ways you could take this and you know maybe one example is like do we have the phone book today like the you know the phone book of Discovery for AI agents and how do we get to I don't know what's after the phone book in this analogy but yeah I'll pose that question um I think for AI agents like AI agents don't necessarily need to be limited to one chain specifically like I the way I see things is that AI agents will just have RPC access to whatever chain they want um so it depends I think on the applications that naturally live where like where liquidity gravitate towards or where applications essentially live on because like the cool point of Agents is that you can do abstraction at the user level where you just give an intent and the agent will act in your behalf so the way thing is not to say like which chain is the best because there's going to be things that are better for specific use cases some use cases is going to require like a 200 millisecond block time like high throughput whatever some other use Cas is going to like do a lot more volume with a lot like stronger assets so you might want to Maxim decentralized thing because the agent is going to give you a higher guarantee that this asset is not going to evaporate in like a millisecond or something um you're going to have that asset live for 10 years on ethereum Main net nobody's going to be able to rug it so the agent is just going to naturally do serious things on ethereum for example um that that's just like my general take on on that that specific question um generally I think like when we talk about scaling compute on on chain like a lot of it is well you probably just want it to run on one computer somewhere you know like if I'm training a model I probably don't want to train it 10 times the problem is like what happens if um like let's say I'm running a long running process and then that particular uh computer that I host that I like rented or whatever dies um then well I've got to like go find another one and I probably want to wake it back up and then like rehydrate it with whatever the last state was and so I think that you can you'll see a lot I think of um like the utilization of public networks and and definitely like I know Jasper's on this with with hyperbolic and and te support and stuff um of saying okay cool well can I like go rent one container go run my compute in it and then if it dies well that's fine I've got some system that's like taking some money off the off the chain in the in the kind of reserved wallet to go start another one and so you you know the the future of this is probably some version of like autonomous life that is like unkillable agents that live forever but but more practically speaking for business um like it would be nice to just be able to run tees and and like stuff inside of a trust execution environment with the the guarantees of it running on like a distributed public network where it's could be getting run by multiple computers um and so generally like could we take uh computation offchain run it inside of te and then um orchestrate that through some kind of distribut Network so we get the guarantees of decentralization as long as somebody is willing to pay you know to to get paid to host that work um you know we can we can kind of uh extend some of the the decentralization guarantees of of the network onto that does that make sense yeah to totally agree this is like uh one research direction that we're exploring we call it hyper intelligence and basically it's like autonomous intelligence that can easily self replicate right like just like what sha mentioned like you can have multiple clones running in Te on multiple machines on hyperbolic Network or any other decentralized computer network and they can even self evolve by S right like they can rent small gpus to find te their capabilities and then redeploy on all these different cons and then they can use uh ethereum or like any decentralized blockchain as the the memory storage uh like as the knowledge spacee so uh although you have say you have 100 clones U handling different jobs at the same time you two can have consistent memory because you have a decentralized network to store the memory and so then in in that sense even 99 out of them are shutting down by the human or like the machines are down you will you will still never die because you still have one clone and still have access to the wallet they can still use crypto to pay and access more compute and deploy more clone of themsel and this way um this will be like the the future better version of autonomous intelligence that a lot of people are dreaming about and um maybe I'd like to add there is another advantage of like decentralization that it's kind of less related to this like futuristic artificial life um um case and more related to like the shorter term like um Digital Services case right so like show was uh referring before to the fact that okay these Registries allow us to have like Marketplace of services I think the important uh proposition here of using decentralized network is actually to minimize rent and value extraction from intermediaries I think this is huge right like you see it in web two there is like apple charges 30% rent uh Facebook like huge advertising profit uh Amazon like same thing um if we have like uh an economy where like there is more and more services going into agents like soon this is going to be like uh big uh uh economic Network and uh if there is no middleman then the revenue can go to the end user who gets like services cheaper that's what we're focused on uh with P but it can also go to like the owners and the developers of these Services right um which is what this agentic Protocols are working on so yeah I'm really excited about uh the centralization uh contribution here in minimizing rent three markets for the one let's go cool well maybe let's try to put our thumb like a little bit more on it so you know on what makes infra good for agentic apps because I think a lot of the things we've discussed here I personally want of my own blockchain right um yeah so so I'm I'm curious like what yeah if anyone could expand on that like what how do we actually make it better for an agent I mean who what bank is gonna open a bank account for an agent right like just creating a private key is all an agent needs to do to get on chain um no existing bank is going to have any sort of process around you know onboarding and creating a trafi style bank account for an agent so it's just way lower friction for an agent to go straight on chain and just start you know using crypto instead of the traditional banking system I think something to be said too is that um in the situations where you can use um the chain like a database you don't have to deploy a database um which uh like right now we're in this kind of like everything is still pretty expensive but in the future like this also just means that like spinning up an app that uh basically is like servicing things off of a network and doesn't require a whole you know crazy user experience to like provide value like like my agent on a service registry um you know it'd be kind of sweet if I could just press like deploy and it's out there and I don't have to manage all of this other infrastructure and services like the I think the the public permissionless database concept at its best like actually just replaces the database um especially for something where you know I there's some questions there about like well how do I SC store secure information but like obviously you know a lot of this is being resolved um and so I think there's definitely like um you know just just practically like you could imagine in a not far future world maybe some more mundane things we just like oh yeah we just like this is how we communicate information to each other or figure out who actually exists like you know um it would be great if I could search for a service and and have that immediately because I you know just it could like lower the barrier to entry a lot for actually developing and deploying apps yeah so uh so basically we need autonomous Finance which is can be done by blockchain we need register and uh like memory which can be done also by blockchain um I think the one very very important thing is compute because compute to agents are like food to human if you don't have compute the agents will die naturally and so hyper body will serve as the compute layer uh to provide the energy and food to the agents and we're trying to build it as a in a very autonomous way so for example like we support crypto payments uh very very early and now every agents are able to just like uh pay in crypto to access the computer that they need and then also we uh every service that we have uh is Agent friendly so basically we provide API for agents to be able to just call the apis to rent the computer they need and deploy um the software that uh they're running their own AI models uh to controls I like Shaw's analogy of databases and it reminds me of one of the Normy ways I've heard I've heard blockchains explained to people which is as unified servers so Robin Hood has their server the NASDAQ has their server Bank of America has their server so on so forth um agents can live on this unified server and I imagine a world where like you go hey chat gbt Reserve uh Reserve this table for me at a restaurant next week and then it hallucinates your dinner reservation and you show up and you're just like where the hell you know like that you like you don't have a reservation all this can be proven on chain on this unified server such that you know wallet a who is the restaurant and wallet B who is me both see okay we have confirmation that yes I have this I have this dinner reservation you have my wallet just in case so you can hold my deposit you know and so on and so forth as we've move further and further up the stack to more and more important you know Financial transactions I see um so would everyone here agree if if they had a magic wand and we wanted to 100x the usability of AI agents today on ethereum is is scaling compute the most important thing we we have to do I mean definitely yes for me I'm open to out of P feedback certainly uh the thing that has gotten in the way of a lot of this technology uh flourishing is just cost I think a lot of the time um you know like why aren't these things better it's like we could actually make them a lot better if we wanted to spend a lot more time and compute on them like you know just like a a data processing task for AI I could have like you know a really complicated pipeline of things that it does but usually I'm like what's the simplest thing that's not going to wreck me too hard on on price price on AI has come down a lot I think obviously price on um which is essentially also compute on on ethereum is coming down because of l2s um and then doing this kind of like well can we provision off chain resources and then communicate directly with them and just kind of broker through the chain um I think the combination of all that stuff is going to lead to like a lot more business opportunities just opening up because they're like actually practical on unit cost um so definitely the cost of these things collectively is probably getting in the way of everything like obviously if ethereum was free to use and compute was free we could do way way way crazier stuff like way like way crazier way crazier um and the only reason we're not is because we're kind of rate constrained by these things so then yeah probably that's the thing that we have to focus on most to enable more uh more applications all right let's say let's say you had it what's the craziest thing we could do it sounds like you got some ideas in mind well I I mean I don't know I I want I love like all the crazy performance art stuff of this I think um one of the the interesting things about decentralization is that we can uh kind of go into a problem space that like a a chachu you know like an open IR anthropic probably would just not ever Venture um and some of that's going to be kind of edgy and very forchan like and some of it could be more like existential um but generally I think that that's good like we should experiment with all these kinds of thorny social issues before um before the AI is too smart and we're like oh wow this is really you know like I'd say we're we're playing with a warm Stone I wouldn't say we're playing with fire it's not really a hot potato yet and like no one's really gotten hurt um and so exploring the solution space uh before we have these problems is probably really good so I I don't know if I could say anything in general but I think like all the edgy like you know kind of NSFW and autonomous agents and like will people have ai girlfriends I think that's all like hilarious and I think we should just run straight out that stuff because it's like probably more harmless than we think and it's also like it's just the the the big companies can't follow you there so there's like definitely an opportunity space there that's unique to decentralized products like like Venice is a good example right is you know uh for he his whole thing now is you can just kind of you know we can do this because they can't stop us yeah maybe I have a slightly different take on why compute is important I mean ux definitely super important but also at some point we need to start adding trust to this agent right like today there is like uh just a token on chain and then like everything else is offchain and like uh if we really want to be able to provide like um services that people can trust we need to add verifiability right and verifiability costs right like uh in the optimistic case maybe it's less expensive but there is things that uh we will need to generate ZK proof especially when these agents operate with data of chain like T have attestation that need to be verified so like verifiability is also expensive uh on the compute side and uh yeah we need to scale for that as well yeah it totally reminds me of a a tweet I don't know if this is your slogan but I I saw you tweet David day uh putting putting the credible in incredible agents uh I think is super important um yeah cool so maybe I will uh open it up for one last question um since we're at time um and what is one idea or product or Vision you think you would like to see hackers building in the agentic ethereum hackathon in the next um you know uh hackathon period um and it sounds like I already know Shaw's answer it's a a 4chan AI agent no I'm just kidding no no I will answer really quickly I would love to see uh the a get into Dows I'd love to see more uh like automation of governance and governance tools uh deep funding and like how we reward contributors or I think things that like are pretty good at AI can solve a lot of that because a lot of it is just like reading chats and seeing who helped who and stuff and I'm very interested in all that stuff I would love to see um AI agents applied to solving the immediate problems for Consumer crypto I'm a big believer that we don't need crazy protocol Innovation to 10x the amount of users on chain today there's just a lot of problems that people face when getting on chain you know like using a wallet it's like not easy to set up um using a block Explorer it's really intimidating and confusing so I think using an agent to help with onboarding could be a really impactful way to bring more people on chain today yeah for me I will be quick uh basically want to see agents become ethereum validators using hyperbolic compute and I think that's doable and we already build some tools for developers to do it kind of took mine but um around Dows as someone who deals a lot with um with Dows but it's really exciting I would say other than that I want to see more like year 2.0 style vaults vaults with agents running like sophisticated defi strategies where they're monitoring the health of alone on a while also you know trading on uh per decks like GMX or or hyper liquid and swapping in and out of tokens um earning earning a safe yield where I can actually actually trust the agent to not run off with the money or go buy the you know the next celebrity coin you I have a few thoughts uh I guess like one that is cool that also D mentioned is the the concept of identity and delegating identity to an AI uh some of the tools that we're building right now is sort of like um how do you know that an agent is acting in behalf of a human um so maybe like just a selfish one is like integrating agent with World ID and then just you know that the agent is uh sort of acting on behalf of a human you can also like rate limit the agent and stuff so there's like cool cool ideas that I have around identity and just selective disclosure of your identity and just authenticating yourself into applications so that maybe like open eye operator or whatnot like these kinds of Agents can interact on your behalf uh in a private way also and authenticate as humans with the application David day any uh wish list hacks for the hackers for me i' really like to see people experimenting with like um wrapping Digital Services and uh that like generates some utility even Beyond like this first generation of agents that essentially connected to Twitter and they just like post some information there there is plenty of like uh data processing heavy uh tasks uh or even not that heavy tasks that like uh this agent can be very good at uh one example uh it comes to mind like someone deployed a bite agent which wraps the door Das API and people were kind of like using it to like uh send some order in natural language sending some intent I think these things so we need to experiment a lot and find the use cases that like uh are kind of like can find the real fit then then we can start building like uh real agentic application on top of them cool fantastic um well yeah I think we uh just are only scratching the surface on a bunch of things that uh hackers can work on here and uh as well as people in the ethereum community to develop to help supercharge AI agents so thanks everyone thank you so much everybody for that amazing discussion and D Mars thank you for doing an amazing job moderating uh we finally got some uh some interesting hot Tis so I'm excited [Music]