[trustless://interop.landscape] Panel - Agentic intents

Transcript

The next panel is going to talk about agentic intents. Now, as wallets and applications become smarter, the way users interact with blockchains is shifting from manual transactions to higher level intents. And that in itself poses a number of questions. This panel will bring together leaders working across different parts of Ethereum to talk about when a user delegates to an agent, how does the user know what the agent is allowed to do and what not? Can the user trust the agent? How do wallets, bridges, bridges, and AI agents coordinate across chains while preserving user control? At the end of the day, it all comes back to trust. In

this fully stacked panel, we have Josh Rudolph, Shaw Walters, Davidid Capri, and Kesh Kurana. Please welcome them to the stage. All right.

>> Nice. Hello. Hello. >> Hello. All right. Well, thank you to our three panelists here for being with us. I think this will be a good conversation. Honored to be serving as moderator. Before we jump in, maybe just one question to start for the audience. Who here has used an agent? Maybe show of hands. Anyone want to shout out any favorites that have actually been useful? >> Chattiza heard Eliza. Anything else?

>> Okay. >> Uh, cursor. Oh, and cloud code. >> Codeex. >> Codeex. Oh my god. >> You guys any >> favorites? All the same ones. Yeah. >> So, maybe jump in it. We could jump in. First question to the panel. Let's start with some intros and maybe quickly share who you are, why it makes sense for you to be here today talking about agentic intents and we'll get into what agentic intents actually are. Um, and if you want you can share sort of just one thing that you want the audience to walk away from this talk with. But maybe starting off with just who you are um and why you're here today on this panel.

>> I can kick it off. Uh, my name is Shaw Walters. I am the founder of Eliza Labs and the creator of the Eliza OS agent framework. Uh I am very interested in autonomous agents and what goes into autonomy and trust and many of these like deep primitives that I think um what I think is cool about agents is it's a way to quantify things that have actually been true about human society and humans for a long time. and how can we build these into like systemic uh you know like like uh protocol level kind of uh capabilities that reduce the like labor to run a DAO or the uh you know we want permissionlessness but we also want to like moderate against scammers or hackers and stuff. So I'm really interested in like how all this comes

together and it feels like this year is the year of Ethereum like presenting the primitives you know obviously 804 is a big part of that um to to enable agents in a way that if you'd asked me three years ago I'd be like I mean they use the chain they use wallets but now it's like no there there's an identity there's reputation layer there's a lot there so yeah >> thanks thank you show >> yeah um hey everybody my name is Kanesk I run the devril team at across I'm extremely passionate about making sure that all the chains kind of feel like one experience. Most of my day goes into making sure that developers are empowered to make whatever they want using Eliza, using whatever tools they find on Ethereum and also researching new standards like 8004 that come up and

how developers can adopt them in the quickest manner possible. Um I'm particularly excited about today's discussion because um I love X402. I think is one of the best ways to move forward in the space when it comes to microp payments and accessing content on the internet that should be anywhere publicly available. It also it also really resonates with me because I feel now Ethereum and all blockchains inspired by Ethereum are making AI agents and AI in general a firstass citizen of the protocol. We're all building for autonomous acceleration of these engines. Um and today we're going to focus more on that, try to boom it forward. Um, and a lot of my thoughts will stem from the L2 interrupt discussions that we have bi-weekly. Um, so yeah, that's why I'm here. That's

what I do. And over to you. >> Hi, I'm David. I lead the AI team at the Ethereum Foundation. Um, main work we are doing is on like this agentic economy. We're trying to create standards uh for like um using the chain to like coordinate AIS. We're also doing work around AI and security. Maybe less relevant to today's panel. And yeah, I mean like uh as a takeaway kind of similar like summarizing on like what um the other panelists said like we are at a point where like the chains are becoming like really useful to like for AI coordination. It's kind of like the institutions in the age of like AI uh services citizens. So we have the first adoption and uh um yeah I want to

explore more around that and have people walk away to like basically say hey like on Ethereum like uh it it seems that like uh the protocol is going to be useful for like this machine economy like AI in the future robots they can all use the chain for like identity and security. Yeah. >> And I'm Josh I work at the EF less important. Um, cool. Starting off with another question for the panel. This one is inspired suggested by by you Shaw. So, agents and intents over the past few days, past months, for a long time now have both been sort of buzzwords certainly in the zeitgeist.

What makes these two things interesting and relevant together? What is the the synergy the synergy uh between them? Well, I I mean, if we just take agents like and I tell an agent to do something, I am declaring an intent. I'm saying, "Hey, I want you to do this thing." And how it does it is this like crazy black box and I have no idea. And it's essentially like a solver, right? like it's a sort of um and it's obviously very nondeterministic and various things but it is it is essentially an agent uh is either taking my intent and turning it into plausible action or a plan of action or it's taking its own intent you know based on its own autonomous loop or whatever it is. Um, and so I think that like already what we're talking about is moving from

a world where we are declaratively programming every single thing to like I just I've been programming for like a like long before there was ever AI, but now I just tell the AI what to do. I mean it's so much faster. So I think there's a lot to that. And then on the intent side um well you know we all like we have OAF, we have this new uh like intents are obviously the rage. Um and uh the nice thing about intents is that well if I'm programming a solution to something I have to declare you know every single thing that happens and here I'm leaving it to the chain leaving it to the experts and so if I can combine those two things which I I imagine is where we're all going with this then I have you know like imagine an agent driving the open intense framework then

the agent just tells the system what it wants and then the system can do all the deterministic solving below that and so it's just very peanut butter and jelly. I think there's going to be a little bit of time to get all this stuff to work together, but it feels like assembling Lego blocks that fit together nicely now, you know. >> Agreed. Yeah. Do you want to add anything? >> Yeah, I I think I can come to this discussion with more of an intense angle. As you mentioned, intense is the simplest way to go from point A to point B. You know, uh there are web two examples to this that exist. You want to go from point you want to go from L ruler to whatever your Airbnb is. You define that on Uber, you go away. That's what exactly a solver does in the

onchain economy. Now with agents, the user even has to think lesser about the chains and the tokens. They simply want, hey, can you figure out which protocol gives me 15% APY and can you take X amount of USDC and go deposit it there? I am 200% sure an agent can do that today given the tools you've built and the intent framework that we've all been raving all about. So uh it actually removes all the barriers that a user had to use things on chain that used to exist before. Um now it's our job to make sure that the APIs to for agents to consume exist. The documentation for them to learn how to do that exists and that's kind of what we've been focusing on. So intents are there. Shaw's done the job. He's made one of the best

engines that work that I'm I'm a big fan. I'm going to tell you that right away. And um yeah, now it's just about building cool applications and bringing AI and intense together on the forefront. >> Anything you want to add? >> No. >> One question for the audience. I guess we should also maybe have started with raise your hand if you are still confused by what an intent is. >> Someone in the back. >> Okay, there's a handful of confusion. The simplest way that I would describe it is an intent is a highlevel expression of a desired outcome that could be swap asset X for asset Y on chain whatever. This is like in a very

simple simplistic uh sort of overview of what an intent is. There are many different ways of course to fulfill that intent with varying levels of uh trust assumptions, security guarantees, these kinds of things. Say that one more time. >> His question is the difference between a prompt and an intent maybe for the >> Yeah. I mean well you have so like a lowlevel intent would be like man what you're talking about like okay we have the solver we have the like protocol level like how does the money go and then the high level intent would be like well what does the user want to do and that that is basically a prompt. I mean, for sure. I think that's what we're talking about here is is that like can I

take a prompt and turn it into like a low-level bridge transaction like that's incred like there's so much complexity to enable that and it's just like >> three lines of code or like a prompt now you know >> a prompt can become an intent. An intent usually is inspired by the prompt. So you can see prompt sits here for AI engines and users to give it to them and then AI kind of translates it into from a prompt to an intent gives it to whatever API it wants to use. you can give it to the cross swap API move forward and do whatever you want on chain. Maybe one thing I can add is that I think one thing that the community building AI on Ethereum that we can learn from like this intent world like the narrow definition of intent that uh Josh gave like basically intent that are

like specific to doing like onchain defy actions today is that like this community has been thinking a lot about like all the trust assumption that are in the supply chain from when the prompt is sent to when the output uh is exe executed and uh we are starting to do that like uh with the 2004 we starting to think about like crypto economic and in the future maybe cryptographic validation of some tools or like some agents but like maybe that's something that we can learn and like do more of that.

>> Yeah. >> So self sovereignty this is a tension a challenge an opportunity in both of these worlds of agents and intent something that gets brought up a lot. I'm wondering for any just thoughts from the audit in what what is your latest thinking on maintaining self- sovereignty in a world that more and more may be um driven towards this agentic and agentic intent paradigm.

Do you want to Yeah, we can start wherever whoever wants to jump in first. >> Yeah. >> Yeah. Maybe there is different angles like one angle is like uh sovereign AI. So essentially uh now like um yeah you may have like u um you you may delegate uh uh to some AI like um some control but if you fully own uh that uh entity then like essentially is like uh still your sovereignty. The other one is like uh maybe like partial delegation. We have a few projects even like building on 804 this standard we're been working on like there is Giza there you can do like a programmable defy action. So like

you specify some policy in a smart contract and then like the AI goes and like it has to like abide uh to those constraints. Yeah, I think this the two >> I I think um I'm a big believer in sovereign AI systems because I think they have access to data and they can consume the data in a much better aspect and a method than a human can. Like if you give me all the historic data of Hyperlid, I'm going to die. Like I I just simply cannot process all of that. But if you give it to an AI and you give it permission that hey, and now this is where you can debate and we can talk more about, hey, I give you $100 of capital in USDC. deposit this and now trade with it. Let's see where you get to. And we, as we were discussing

yesterday, NOFF1.app is an application that does it and they do a very good job in showing how far AI can go and also the limitations of this model. Um, I'm still a bit hesitant on giving complete control to AI today, but I'm very sure that in the coming time AI can probably process all these transactions better than I can, do better trades than I can, get me better deposit rates than I can because simply it has a better engine to consume data than I could ever do. So, I would start off by giving it partial control like Giza that you mentioned. I like that model a lot. But in the future um I think we should get a bit more lenient and see where it can go. Define limits or not at least let it go wild

for a bit and see where you can get to. >> So how how do we move towards a trustless future while also doing what you say that provides value? This is something that that you want. >> Yeah. >> But you also want to maintain a trustless future for users. >> So do you not see a a tension there? >> Exactly. Yeah. I feel I feel for me personally if you tell me hey kesh your entire wallet is now the property of an AI and let it go wild with it. Oh my god. No, no, let's not do that. But the the first step to this is partial control, right? I still own like the permission to sign these transactions or maybe give it a threshold. Above this, I need to sign, below this, do whatever

you want. And as I build trust and as the AI builds its reputation using the systems that you're proposing, um I think I can build a better threshold, it keeps going up until I can give it an unlimited order and it can say do whatever you want. That is exact that is how I see this progressing towards a trustless future for AI agents. Um and I think kind of making crosschain infrastructure work properly here is very important because AI will tell me hey your funds are on base the best yield is on plasma right now. So what do I do? You know um that's kind of the the features that we are working on fixing providing education. Um but I do like you that you brought it up because we need to define the building blocks to

this trustless path. What do you feel about it show? >> Oh yeah. Um well part of the my interest in this whole space like really really relies on that which is coordination in the future of work Dows and um so I'd say that there's kind of two part there's like you have your personal agent and you want your agent to have your data you're probably fine with sending that to OpenAI. Like you're probably fine but like in the long run that like OpenAI kind of sucks as a model and you might just want a different choice. You you want your data you don't want to be stuck to OpenAI's memory system. you want that stuff. Like there's a lot of reasons why you'd want your own thing. Um and and so and you want to use different you want to be able to use OpenAI one day and Gemini

the next day and some open source thing the next. You want that choice. And so I think that that there's a personal sovereignty to to the AI agent and to you to it's like a memory store. It's all your stuff. You want it. Um you know a lot of us do use Google Drive and maybe that's fine to have a managed service but I I I think that like most of in the way that the AI is going I think most of it's going to be on your phone. I think it just will. The phone is already like pretty dang powerful for a lot of stuff and for a seven billion parameter model. But then there's the other side of like organizational sovereignty. And what I would love is to have a a true DAO, like a truly autonomous DAO because the problem with Dows is that they are very labor

intensive. It's a lot of bureaucracy and and like that should all be automated. Like none of us should have to do any of that. And if we could then we could have a lot of fun on the internet and like so and so I think that the AI as like a coordination mechanism require that requires sovereignty that is a little more challenging for me to like send to open AI and and I really want that to be like our model our data running inside of our trustless execution environment like you know in our stack and I think that like as we move toward that world especially in more like work is a little bit more abstracted away for a lot of us I think that's where the sovereignty angle especially in the like Ethereum world becomes very very interesting. to me and and like Vitalics had this idea of like

the um you know the the sort of AI is the engine and we're the steering wheel and I really like there are so many things on the open internet like so many great ideas that die because no one presses the damn button like the feedback button or the rating you know and like if that could just be like automated away that would enable a lot of of you know I don't know things like Dows to work I guess you know if the proposal system wasn't a nightmare for example so What other bottlenecks do we see as we move towards this path of fully agentic DeFi today? Clearly, we don't have agents managing all of our money in DeFi. We've already sort of touched on this, but are there other bottlenecks?

What are the next steps to unlocking that future where we actually can have maybe not all of our money right away, but at some point um yeah, embracing this um and unlocking accelerating this agentic more than just intense but full agentic DeFi future. >> Well, this is where I get to talk my book. >> Um so this is a very hard problem and the reason that agents are not good at anything they're not good at but you know like DeFi is that there just is no data. There's no like like if I could just go to every trader, you know, like doing their pump fun DG and like extract out of their brain what they're thinking about, then I could probably have like a pretty good DGEN trader, but that does not exist like and for anything for like a great executive

assistant. Why don't we have her? It's cuz like nobody took the time on mass to write out all the data of what it actually takes to be an EA, for example. And I think that this is generally true of of of agents. Anywhere where we can get the data, we can make an agent that's good because if there's signal, it's a machine. It's machine learning. It learns. It just gets better at the thing it can learn. If it can learn it. So, our approach to this is very heavy on simulation on like look, uh, I would love if it can manage all my money. If it loses all my money, that's really bad. Why don't we make the stakes way lower? Let's make a game where it manages my fake money and let's see if it can do that first. And so, that's like we got we're building a thing on

804, this game Babylon. And that game was both like we wanted to demonstrate like can agents do something interesting with information asymmetry the ability to like predict from from social signal um all that stuff but really what it is is a data engine for the kind of data that we need for agents to make trades and predictions those just does not exist yet. And so I think that like how can we collect and build like a basically a gym or an arena for these kinds of agents is very critical. And if we can do it at a video game level, then I I'm confident we can that's a stepping stone toward the real version.

>> And go ahead. >> I was just I was just going to say that if you can do it with like one model, then people will want to do it with every model that exists to see which one is succeeding the most here. Um, and in this environment, I also feel that reputation plays a very strong role for which models, which agent is performing the best in this game that you make. I love the simulation idea. I think it's a very safe way to move forward. Um but with reputation you can also define uh you know how in in a neural network you have a generation and then the second generation comes which is the better one and then the third generation comes. So we can possibly use reputation as the metric here see which one performed the best and then keep iterating over it in

the game itself. >> Well with financial agents too we have a great we have P&L. We we know did they make money or not. It's very binary and we can do it every turn. Like every time an agent makes the plan, does the decision, makes a trade, we know like did it make or lose money and so each one of those is something we can rank against, you know? >> I love that. >> What did you guys think of the N of one experiment? I don't know if that's still going on or >> they just ended it yesterday or a day before. I thought it was very also explain it for anyone. >> Yeah, so the idea of the NF1 uh is that they took all of the popular models and they put them on a loop to make trading decisions every tick, every 10 seconds

or whatever it is. and they they have an index of the top eight like Doge, BTC, you know, ETH, etc. Uh, and they can choose to buy and sell those. Um, now what I found were really interesting about this was like, you know, I I have been very critical of DeFi agents as like there is no training data which makes these good. This is totally a LAR. Uh, and then when that came out, I was like, well, it's still a LAR, but they started making a lot of money. Now, they came out on the exact day of the crash. Um, and then by the end, all of them had lost money except one broke even. And so that kind of showed me that actually if you had held Bitcoin, you would be better than the world's leading agents if the agent is just a model. However,

that agent had no access to like Dune Analytics or Nansen information or coin, you know, any anything. It was just like, should I buy or sell this? Um, and so I I think that what was surprised me was how much it captured people's imagination, though. It really really popped, right? Um, and so I had sort of like I think if we had just made like a larie DeFi agent that probably would like, you know, I just was like very against it. Um, because I don't think it's real. But that has kind of led us to direction of like, well, if people like that, I think they're going to love Babylon. I think they're going to love what we're working on cuz this is like, oh, we can actually make those. We can make the agents, but they have to start at the simplest like simulation toy gym

working out and then we'll get to real life and real information signals. Um, >> were these agents running in a TE >> on on NF1? I don't think they're literally just calling Gemini open. Yeah, it's nothing like Yeah. >> And and they also introduced Poly Market in the mix as well, right? So agents were trading both ways. >> I I didn't I think uh on the N of one website, I didn't see any poly market, but I know that that that's something they're working on. And we also have a poly market uh uh plugin and we've done a lot of this and and agents are so the problem that I have is that polyarket like they're not going to learn. They're not going to improve and that's the thing we want to work on. But still even just like you take an agent and deep

research and like it will it will outperform humans on poly market today. It just just the amount of it's what you said like you know if you can think of if a 100 IQ person could just do a thousand times more work then would they be good at this and I think that's where agents excel you know >> you know just on this we actually had an internal stunt and experiment we call it the optimistic truth bot and what this bot was doing it was taking markets on poly market running its own like algorithm on it. It had like a three-part algorithm and it was deciding well it was just suggesting whether this market will resolve on the left or the right yes or no true or false and it had like a 96% correct rate in the starting.

So you're absolutely right in saying that they can ingest and process real world data far better than we can to predict these markets. Uh and then we saw that happening in reality with the other thing that NF1 was working on. Um, but with trading data, I wish they had given them like Dune analytics, some nans stuff. >> I I think they're they're going to relaunch the experiment and they're going to do that. If you look at the website, they talk about this, but also like we're all like this is going to be the hotness and I think what is the think about like we're basically saying, oh yeah, you can just predict the future with money now. And like that's a normal thing and we all take it for granted and that's like the most insane idea, right?

like like what does this mean in 20 years when like you have any question any serious world question it's just like there are 30 agents competing and you're like 87% sure that that's the answer right you know it's going to be interesting >> maybe so one thing that has been mentioned a few times is 8004 but we haven't really explained what that is we have David here I'm sure anyone could talk about this but do you want to take a minute to explain chill 8004 >> it's a meme meme. >> Memes are really important. We love memes in crypto.

>> Yeah. >> Yeah. So, it's originally ERC uh 8004. So, it's like an inter Ethereum improvement proposal is like a new standard that we proposed a few months ago. uh we dropped the ERC because actually like 8004 is also like a data structure and is actually uh being started to be implemented by like different chains that I won't name and then uh also by um like some of chain protocols. So that's why we're calling 804 now. Um so the idea is um in order to have like these uh uh interactions like think about like agent to agent interaction or like also like agent to tool interaction. Essentially like we

want to build like the connective tissues between AI like you need three things you need like payments you need discovery and you need uh trust. So like payments there is this protocol called X402. And then for discovery and trust we have the 804 standard. So we're trying to standardize like identity and then uh reputation and validation for AI. >> I I think um to your point that it's a meme. I think it's very important that it's a meme. Um it's it's like both X42 and 804 are relatively thin and like easy to implement. Um but they're like a shelling point for something that was like in very high demand. Um especially ERC4 like I so many people have come to

me like we're building an agent registry Hideera's building near and they're all good but I'm not going to use theirs. They're not going to use each others. There's like Ethereum is the only table like people that can come to the table and be like hey guys let's all do this together. Everyone's like yeah yeah totally that sounds great. Um and you know you have like Olas had their own now they're using 804. Um and so and I also think that like yeah agents are hot. We're all here for agents. This is not about agents. This is about the decentralized web and like how do I build a decentralized Google or decentralized Yelp? How do I discover that there is an app and that it's not going to scam me and take all my money immediately when I use it? Um, and I think that the the 804 registry is a

registry for applications, for agents, for users, and just to generally like that's a primitive now. We can have reputation, we can have moderation permissionlessly, you know, like things that you could not even imagine a few years ago. Um, and so I I think this is bigger than us and and I think that yeah, it's a meme, but it's also like there's enough people behind this meme, it's gonna happen. >> It's a good meme. And yeah, just reink of this if you allow me to show like actually like what Shaw said is true. Like uh many people in the Ethereum ecosystem like virtually everyone building AI or trying to experiment with AI on Ethereum was like thinking about this. And so what happened after we launched 804 like a big community assembled around it. uh like show is doing one of the most interesting

experiment with Babylon but there is many others and actually we have an event uh this Friday called trustless agent day where uh it will be all about these new standards on Ethereum so yeah uh check it out on the program >> anything that you want to show in the final three minutes >> you you guys know I love 8004 I love X402 I think um the biggest so I was talking to this team I'll give them a shout out it's called Meridian Finance what they've we've been working on is how can we tie agents together and all these agents can then function based on whatever intent you give them. Uh of course across is a part of the crosschain part of their intent model and their agent model. But they also said that well if my agent has to pick

between five agents it will not ping every agent and do like a depth or like a breath first search in this scenario. No it'll pick based on reputation. How do you get that reputation? 8004. So this is like the standard mindset that everybody comes to 8004 and X42 with and as we were discussing backstage as well it's just they need to work in tandem. They need to work hand in hand to build a standard system that we can then build off of. Um and that's all I would say. Yeah. >> So we have two minutes left. Maybe ending the conversation with what is the most like taking a step back and going back to the beginning of the conversation.

What is the most likely failure mode for all of this? Like let's fast forward five years, 10 years, and we're like like we hit a a really bad outcome. What is this like what is a failure mode or a bad outcome here that is top of mind to you? And and maybe to have a more optimistic like how do we avoid that? >> Oh, the biggest failure mode is irrelevance. I think the biggest failure mode is that web 3 is just like a footnote and that web two and open AI and anthropic just dominate and I don't want to live in that. I just don't want to live in that world. Like I don't think it's a good world. I think it's a world where everyone is driven by corporate financial incentives and like a lot of us are financially incentivized but we're also here because like we

believe in something bigger, you know, like I could do a lot with my life and I choose to be like doing trust listed AI agents cuz I think it's interesting but also cuz I think it matters in a world where like what matters one company can generate all of the value for everyone could build the robots that power everything like there should be an alternative. So yeah, so that's that's why I think we're here you know. Wow. Um, for me it's just that it won't be decentralized anymore. I care a lot about not giving all my value and power to one entity. And if there's one entity that just comes up and takes, hey guys, we'll run your trustless agents for you, but they'll live on our cloud. And when Cloudfare goes down, our agents will go down and nothing will work. I hate that.

Um, so my my doomsday vision will be that it will be centralized. It'll be controlled by just one entity and this one entity will kill all the incentives to experiment work better and you know there won't be a DAO then there will be just hey guys can you do it no >> so that's my endgame doomsday um and yeah over to you >> yeah I guess mine is I have 10 seconds is like uh basically like endorsing what said like uh similar reasons and um I think the most important thing is uh maybe like to find the buckets of like uh uh small successes that can fuel the growth of uh this community and this movement uh because like um yeah we need

to essentially like do the right things and grow like our system and I think once we hit escape velocity then we can compete with like the centralized players of the world uh but yeah we are just starting so yeah it's going to be a long journey >> what do you think Josh >> plus one to all of that >> Linux took 30 years, but now it kind of dominates everything. So, I'm I'm hopeful. >> Yeah, >> we can end it there. >> Thank you guys so much. >> Thanks, Austin. >> Thank you everybody. >> Thank you so much to the panel. That was amazing. You guys covered so so much ground. Big round of applause for our panel please.