ETH Denver 2025 AI Summit The full stack - What it takes to build a good agent

[Music] okay oh are you Tim I'm Tim oh amazing cool well I'm glad I'm glad you're able to join us um hi everybody well I'm really glad to have you all here today for this AI agent panel we'll be discussing the full stack what it really takes to build a good Agent um I'm particularly really excited to be here I lead developer relations at cell and we focus a lot on help Builders create AI agents and there's a lot of different trade-offs when it comes to building AI agents when it comes from uh security the tech sack the trust layer and different assumptions we're making so I'm glad to have this panel today um with all of you where we can dissect those and actually begin to do some hot takes some spicy takes uh about AI agents and uh how to build them uh so to get started can you each introduce yourself and a really really quick like just one liner on what your company does um and why what you're working on is important for AI and AI agents who wants to start us off I'll start hi I'm Tim cotton uh I'm from scripted we are building uh layer one for AI agents called enori Network and we build AI agents in general and if uh we have an upcoming stack called AVB for autonomous virtual beings and fun thing about me my first job was programming AI for video games in the early 2000s at EA awesome um hi everyone my name is Alonso I'm one of the founders at Crosman we do web two like apis for web 3 functionality one of our biggest focuses right now is um helping bring AI agents on chain and on that regard we have hosted apis for things like tokenization payments and wallets that work well for agents and we also have a few open sour projects uh such as go SDK and SDK where there's 200 different onchain functions that you can directly integrate in Frameworks like Eliza L chain versel AI Etc and as a fun fact in a prior life I was going against agents I created the infamous I'm not a robot when I was at Google so apologize for that trying to redeem myself now hi everyone uh I'm Sebastian Co at memberi uh we're at def5 oh is this this better right thank you um I'm Sebastian uh Co at Ember AI uh we are a defi Knowledge Graph for agent execution um I'm working on uh kind of a similar problem to you uh I think we have a very shared approach to how we want abstractions to help people built these agents um spearheading the arbitrum agent kit uh in initiative and uh fun fact about me I was in the magic gathering World Cup when I was 15 hi I'm Nia I'm the founder of perah um perah makes it easy to uh provision uh embedded wallets uh as well as kind of manage uh liquidity across applications so we build a bunch of developer tools um recently uh related to the topic of this panel um we launched an Eliza OS plug-in uh and a fun fact about me uh I was uh involved with uh NLP in the uh pre-m days and so the um the nice hardcoded uh neural net uh world and so it's amazing to see where everything's gone awesome well thank you all for introducing yourself uh to kick off our panel today I wanted to ask our first question which is why do AI agents need blockchain at all uh what can they do on chain that they can't do offchain he wants to kick us off yeah go ahead um I can take that one so to me I think what I would say is to me I think AI agents will choose at some point what tools they use and I think they will choose to use blockchain in many cases when they see that it's lower friction to do things such as make a payment maybe lower fees and faster execution um to acquire an asset if it's tokenized to get yield things like that so I think to some extent the way I would say is more like um it's not do they need crypto it's like if we as a crypto industry are able to lower the barriers of adoption we have a better technology for moving money and value around and they will come and use it they are more they're going to become more int elligence than humans and they're just going to care about uh throat P reliability fees things like that uh I I'm just going to jump in from the uh hopefully the scaling perspective a little bit um just I'll try to keep the story very short but when I worked on the first generation of massively multiplayer online games we had AIS we called them non-player characters we didn't have deep learning accessible but we did have lots of Behavioral scripting and those things were expensive to run on the CPUs we could not resolve stage for hundred of hundreds of thousands of them running at the same time in the same world the players were cheap players are just movement you know players send little packets to the from the network stack AI agents required tons of computation decentralization should promise US has promised us that we should be able to find a way to Shard those computations actually scale up Ai and that's exciting to me and I agree with you by the way that in that case if we have parts of AIS running in a decentralized context well sure blockchain is a natural payment and accessibility layer for them yeah I think that's really important and we talk a lot about uh the importance in blockchain space of decentralization and the other piece that is huge is trust right a lot of us work in this industry um because of trust assumptions and I think that's really fascinating when it comes to AI agents um specifically I wanted to talk about Ember AI that does a lot of stuff in the defi space and when it comes to defi there are a lot of trust assumptions that you need to make um as we make uh AI agents fully more autonomous how do we ensure that they actually stay trustworthy uh specifically are we just shifting our trust problems from institutions into unpredictable algorithms and I would love to get your perspective on that session yeah thank you so I think there's many layers to the question of trust uh I think there's one part which is trusting your financial assets to an agent I think there's another part which is trusting the agent execution and the logic behind that I think there's another part which is trusting the actual inference mechanisms and the actual model execution these all have different layers of trust but I think at a high level what's important is when people are trying to engage with AI agents and kind of accept this um nonhuman perspective on whatever it is that they want to do there needs to be a clear uh way to andark or frame how exactly it is that this agent will be doing this stuff for you right so observability sitting at the core of all of these things is probably the most important for generalized trust for you know all the layers of complexity yeah um curious if anyone else wants to comment on that I I just I'll just ask the crowd show of hands anyone play with Salon coins recently yeah okay um if anyone jumped into like radium uh concentrated liquidity Market making you know the clam pools how great would it have been if we had an autonomous agent that we could have given G our funds to to jump in and out of clams for us and get out when volatility was going nuts and also make sure that we could trust that agent that it wasn't just siphoning off our funds now that that's where I like this idea of credible neutrality provable fairness yeah 100% yeah I mean especially when it comes down to how do we make sure AI agents don't just go Rogue with our funds uh I'm really interested to hear from natia and uh what you're working on at perah what specifically would you advice you would have for Builders creating AI agents to ensure that they don't actually go Rogue yeah great question um I think Ai and crypto are such a uh kind of odd perfect pairing in the sense that uh llms are most infinitely non-deterministic um and you have that coupled with uh crypto rails which are irreversible and deterministic uh and when you bring these two things together um there's almost this interface where the two meet uh where you need to kind of patch these two very different realities and um one thing that we think a lot about is how can you um allow a agents to reach a goal state or to um maximize with respect to some objective uh but with some guard rails to ensure that um maybe there uh if there are undesired steps uh that those are either prevented or if there are desired steps those are um kind of maximized like one example might be uh certain protocols that um maybe recently had uh a hack on them um you might want an a agent to avoid kind of leveraging um a a compromised liquidity pool um a a multi- that has um that has kind of a known vulnerability and so uh we think a lot about like how can you set guard rails as well um so that you get to the right destination but also uh do that kind of safely or leverage kind of human knowledge to to get there yeah I like what you talked about around guard rails for AI agents and I I think that's an aspect that makes building on chain on chain AI agents very exciting you can actually program different things around how uh money is spent or actions they're interacting on chain um Alonso I know uh Crossman is also uh the team behind goat uh can you tell us a little bit more about goat and what makes this framework unique compared to maybe other AI agent Frameworks okay um uh so go SDK not be confused with the animal or the token is an open- Source Library the biggest open source library of unchain actions um available directly to consume by AI agent Frameworks so the idea how this works is there's more than 50 companies that put their their wallets their like um onchain protocols whatever it is their their like data feeds that they provide for onchain tooling they all put them there as one library and then the framework takes care of ensuring that from a plugin there uh agent Frameworks like Ela Versi L chain Etc all are able to install those plugins without having to write any code um I think what it makes so one thing I would say is like to make it different from things like Elisa and things like that Elisa is an agent framework or virtuals or serery that more like run the agentic loop this is a tool kit it's more like a collection of tools so they're compatible there's other tool KS out there and I think at the end of the day what matters the most is which one has more tools and higher quality because they're all open source many I think will eventually just merge because it's all free open source they really not like a competition um but yeah encourage you all to to check it out and see and things that we're doing there that I think are potentially relevant for the topics of security and so on um I think a big problem we're going to have is supply chain attacks on AI agents uh so something to some extent similar to the hack we see uh we saw in um byet although that was on the front end but imagine your AI agent is doing a swap and it's just importing some code from a random repository um you need to make sure that the code that they're running is thrusted so we're we're going to be implementing some things there for formal verifiability that you don't have like uh or or not guarantees but at least get as close as we can to make sure that the plugins there don't have like a drainer or things like that that but uh yeah I think this we're still at 1% of where we all need to be as an industry so we'll see how things AV very fascinating I mean you talked a lot about uh a bunch of different things we can do for onchain actions we talked about programmable wallets we talked about I know a knowledge graphs um and with all that we our panel is about the tech stack behind AI agents and so I'm really curious for all of you what emerging Trends or Technologies do you see the future of of a um agent development yay this is this is the fun part for me um I I just want to follow off what Alonso was talking about because he called it out clearly with Eliza zery and others um there are multiple open source Frameworks right now that have just gone parabolic uh on GitHub and if you look at like what Eliza Labs did with Shaw's project um they managed to collect hundreds of us to contribute to this repository and create tons of endpoints for services everything from voice generation knowledge graphs all of it right and multiple other Frameworks looking for their own edges did kind of the same thing Pippen went a little bit differently um Yohi decided I have a different kind of loop that I want to do my value is in the way that it autonomously creates its own functions so I'm just going to go hook up to a provider not Crossman in this case but a provider of these services and it can just register and find those services so we're finding now that the framework works that move towards learning memory and actually more complicated like thinking processes in planning and strategies that's your V2 that's what's actually going to be important over the next couple of months whereas we're going to start depending more on services like over on this side of the panel that's awesome well thanks for sharing I know we only have a few minutes left and I wanted to do a quick lightning round uh for a few questions I have um starting off and I would love to have each of you share so you could just like yell out that word um for you uh what is a really interesting AI agent use case that you are most excited about who wants to kick us off alien life forms alien life forms fascinating onchain alien forms fasc onchain aliens I want um executive assistant in telegram that manages all my chats for me yeah I think the financial inclusion element of Defi and specifically the the lowering of the friction with AI agents and llm interfaces is something that it'll take a little bit for us to get there but as you see you know crypto penetration in like Emerging Markets you you'll also I think see a lot of that um you know being able to enter what a financial Market might be for someone in the first world for people who don't necessarily have uh the expertise or the time to dedicate to something like that mine is uh also personal assistant um for kind of very Niche uh Financial tasks like calling up a number uh when people don't have credit cards enabled uh like those really kind of laborious like phone call involved uh tasks I like that I can't wait till my personal AI agent assistant exists and they can automate My Life um and then my last question I have for you and maybe any builders who end up watching this later on what's one piece of advice you would give to a builder who's getting started building AI agents well right now you have multiple Frameworks to play with so just download one edit a character profile run it and then learn exactly why you don't like it or if you do great but if you feel like the character profile just isn't enough and you can't hack it with just a prompt now you have something of value to add to the industry that would be great yeah to to latch on to that I think the idea like let's not forget that llm calls are all about context and context management so however it is that you're feeding content into the queries that you're making just really pay attention to that because that's probably going to be the source of all of the errors that you might run into mine is uh you can just do things uh and then also uh keeping in mind uh what value do you want to add uh What uh incremental kind of extra functionality you want to provide and for who like always thinking about the user uh mine would be don't launch a token make make an agent that actually provides utility fascinating thank you all so much for joining us and uh sharing your knowledge uh on this panel um really enjoyed having you and yeah definitely make sure to go check out all of these speakers and their products um as well so thank you so much thank you thank you shout out to the beans at AVB and everyone else who Ling