AI Agents Will Replace Dapps

Transcript

Hey everybody. Uh so thank you so much for coming. Uh so um So we're Coinflow. Uh we are the very first uh AI agent that is capable of research, execution, and on-chain automation uh of any of any on-chain action. So uh today I'm going to talk to you about why AI agents are going to completely replace the dApp layer, uh which might be a little bit of a controversial thesis, um but I I hope I uh inspire you uh for the the future of Ethereum.

Uh so a little bit of background for myself. Um I previously served as as head of product and then lead of operations for MetaMask. Um I am uh the inventor of the the wallet business model uh for for growing and monetizing the MetaMask wallet. And uh this product and uh our our thinking uh for for crypto user experience is the product of seeing uh what the challenges were uh in in building uh both MetaMask and scaling Ethereum to uh broader audiences.

Part of our thesis is that the the dApp model for how people use Ethereum and how people use crypto is is a fundamentally saturated model. Uh that that what we've seen in 2022, that was essentially the total addressable market of the people who will be able to use Ethereum in the dApp way of using a browser extension, injecting JavaScript into websites. That model uh we've we've basically seen it. Um We If we go back 10 years and we think about the use cases that that were trying to build on Ethereum, they are very diverse. They were uh they very interesting and if you go down the list from from Devcon 10 years ago, I think

you'll find that almost none of these none of these user experiences or use cases have have delivered what we as as Ethereans believe that they would be. And we look at how people are using AI and we have like punk funds Mayhem mode for faking volume on tokens or CZ tweeting about world's greatest casino and you know, we look at where the the adoption curve is and so that this is prices compared to users. So we see user adoption declining. We see

asset prices increasing although they're very decreased from the time that this slide was created. And and so this is a really alarming adoption curve that looks sort of like the Linux adoption curve. And the Linux adoption curve was an adoption curve that one hegemony all of you know, the entire web and the internet they run on Linux, but we never succeeded in breaking into the user layer. We seated the user layer to proprietary centralized tech and and so when we have built this product, our our goal is to shift things back to to the internet's adoption curve and one that can maintain the cypherpunk values of Ethereum.

So why is this possible now? So so Coinfellow has brought together a few different technologies. One we have 7702 in the Petra upgrade that allow us to upgrade EOA accounts to smart accounts. That means that the existing liquidity and the main problem of inertia for for creating a new user layer is actually upgradable and we can now get permissions through two EIPs 7715 and 7710 to get delegations and permissions from the users existing wallet to automate certain things or to use assets in ways that the user has tight control over to delegate permission to to a to an agent. And that has been brought

together with 8004 and 8083 to create things like registries of AI agents, registries of verified contracts and we have models that are frankly fantastic and that are are way better than what we had even three months ago. So we've built a product that we've launched here at ECC that is the very first general purpose AI agent for doing any on-chain action. It gives you a self-sovereign agentic environment that allows you to call any contract, any on-chain action and to to do so through an interface that looks very familiar to

to a mainstream audience. It is an experience that is mobile first. The DAP layer never translated to mobile. If you were to look at the stats of products like MetaMask or other browser extension based wallets, these products work very well in desktop environments, especially for DeFi professionals, but the the translation of using a a mobile interface for accessing DAPs has fundamentally just not happened. DAPs do not work well on mobile. It's not a problem of let's make the mobile wallet slightly better.

We need completely new agentic interfaces for doing this. So so Coin Follow is a product you see here on the right. We provide you a chat based interface. You can do things like say a simple prompt like let's swap some USDC for ETH. When you do that, it is not just a raw chat experience. It is one that gives you uh that that gives you familiar interfaces like a swap UI element or other other UI elements that are relevant to the action that you're taking but that stay within the context of your your mobile app or your web app. And you can access any

DeFi contract. You can do any on-chain action. You can access any DeFi protocol. Anything that you can do with a dapp, you can do through a single chat interface. It completely replaces the dapp front end and it can both carry out execution of individual actions or it can automate specific actions. So, uh so we have a an example here. You can you can by the way we have a open claw or any agent skill that you can use with open claw or clawed or whatever product that you're using it with. And you can define uh specific actions. So, we have a range of them here.

So, the agent is making a proactive suggestion that the USDC Ave position is approaching liquidation. Do you want to add liquidity? And then the user is is adding is approving a transaction to add 200 USDC to to that Ave position. Or or it can also you have an agent that can understand context that can break things down for you. And so, a user could say something like sell all my meme coins to buy more ETH. And the agent then tries to figure out what are the meme coins in this user's wallet that they want to get out of meme coin positions because meme coins are declining. So, it comes back and it says, "Your I think your meme coins are SHIB,

Pepe, Brett, and Pengu. Do you want me to sell these four?" The user approves that and an on-chain action happens. This is a much more accessible way and it and it also means that you can bring your own front end. So, we have tools for developers to either you can consume our agent as an 8004 agent or you can bring your own front end using our CLI product and this is available today. Not every single one of these automations is available today, but they will be very very soon.

So, to make this happen, there's there's two things that make this type of product possible. One, we're we're both getting delegations from your existing wallet using 7710 and 7715. We also deploy for users that don't already have a wallet, you can deploy the MetaMask smart account kit. This is an embeddable smart account. It's a new product from MetaMask and you can deploy this directly from within our product and you can actually deploy this this smart wallet and then use that smart wallet for ongoing transactions as well. And then the second part that is really important and is that the agent needs to run in a trusted execution environment that is

self-sovereign. It is not enough I mean there there are different layers of security. We need to secure what the agent is allowed to do on chain through those two EIPs that I I mentioned, but we also need to make sure that the model that is being run is the model that the user consented to run and that the the agentic environment is one that is controlled by the user. We don't want to create a new layer of custodial and centralization that that represents a a security risk to the user.

So, this is a little bit what um an agentic onboarding journey looks like. Uh so, in in our approach we have the user sign up either with their existing wallet or deploy a new smart wallet through SSO credentials. If you have brought an existing wallet, the agent can get contacts on your wallet. So, we can start making what we call smart move suggestions. So, it can mention that you have certain assets that you're not earning yield on or where a place where you could earn greater yield. And you can use this for discovery of new opportunities. Our our system prompts are increasingly uh being developed in a way where we we consider this like a a rabbit hole deepening

protocol. So, the user has maybe done a small subset of what's possible in DeFi today. The agent can look at that and start making suggestions about uh new ways that you could be using your assets um or uh take you deeper into your DeFi journey. And then any action that you want to approve uh or or that you that you propose, you're going to be able to review that in a a human decipherable way. The agent presents you a plan and then you approve that uh similar to to what many people are familiar with today. Uh we we do have some users who ask for like a full degen mode without approvals um and and maybe that'll be uh behind a toggle or something someday.

But um this architecture requires multiple agents, multiple models in a uh in a in an architecture where different agents are are reviewing and simulating transactions and uh proposing them. Uh the agent needs to have DeFi native context and memory. Um it needs to learn from uh uh your your behaviors and improve those over time. Uh it needs specific evals that make uh on-chain actions possible. So, you can you can see those through our open claw skill or our agent skill, I should say, which is available. And then the models themselves and where we're going is that the large models

that we see today, they need to be pruned for use case specific things. We don't need the the history of of cooking in in our AI models. We need smaller models that are specifically for on-chain actions and then those models to be improved with Laura and Dora data sets. This also points to, you know, Vitalik was was speaking about the AI hegemon problem at Devconnect in at the end of last year. We're we're seeing increasingly things moving towards personal agents and this is the embodiment of Cypherpunk values applied to to AI agents. And so, we we do see a future where people will be able to

completely run their own agents. We're not there today, but we are quickly getting there. And when we combine delegations and permissions through the 7710 and 7715 model, then we get to a place where what the agent is able to do is only the things that the user has decided that their agent can do. And I have a typo there. 8004 agents. But we are getting to a much better place. And then from the developer side, we have the So, a user can or sorry, CoinJour or or another party that's developing an agent, you should

register your agents in 8004. Our agent will consume your agent. We're not going to be the best. We're a general purpose agent at every action you want to do. So, our agent will consume other agents that do a copy trading bot or or any any use case of the that other people are building for their agents and then our agent aggregates those or your agent can also call our agent and use it for its own skills. So we are available today. We launched this publicly two days ago.

So please go and give Coin Fellow a try. It's coinfellow.com. It is still it is still an alpha build and there are many things we are rapidly rapidly improving but we want your feedback. We want to work together with people. We're working with a number of wallets, DeFi protocols, other people like that who are also embedding and using our agent within their products. So this is a really exciting thing for us to to work together with the developer ecosystem and also as a as an end user you can use this directly yourself. You can use it on mobile. It solves a lot of the UX problems on mobile.

Yeah. And happy to take any questions. Let's let's think. Yes, you first. Hi, you said during the presentation that it was deterministic and auditable. AI is notoriously non-deterministic. What steps have you taken to make sure that that is true? Yeah, so that is so that is our where we are getting to. We are working with trusted execution environment providers.

I don't want to name the specific one at the moment but we've been working with people that are doing like you may have read about like Eigen AI's products for example that provide a verifiable AI environment that operates very differently than a a LLM running in a hosted environment. There are others that are also doing that and so we're we're working with multiple partners to try to figure out the best way to to solve that problem. That is not where we are today, but we do believe that that is essential to where we're going.

And then there was another question here and then here. Yes, yes sir. Kind of a follow-up to the first question. What risks are you introducing by by doing this layering and are there any risks of errors? Yes. Yeah so it is a radically different security model. And so if we look at the status quo of where we are today versus what Coin Flow introduces or or other people who are building similar products, we do believe that this is profoundly more secure than what people are experiencing today. Like while we're here at ECC, Steakhouse, the

front end was hacked. It points at malicious contracts. That seems to happen on a monthly basis through DNS hacks, through fishing exploits and those sort of things. So there there are new risks in the agentic world where you know, you have the risk of an AI agent hallucinating or doing something that you don't want it to do. That is our entire product work has been focused on solving that problem. That's why we've adopted this this multi-agent and multi multi-model architecture and tried to to do everything that we can to simulate transactions and to have user review and approval of transactions

before they happen. There's also other attack surfaces like you know, people trying to train train models to to intentionally hallucinate and we're we're building mechanisms for those as well. But basically what we've done is created an entire layer of different security things that make doing this far safer. It is an early alpha, so you know, today, you know, it's it's totally fine to expose a small amount of funds and experiment and break things and we'll have learnings.

But generally speaking, today we have very very few issues that we ever see of hallucinations. And and we've been we've been working on this product for for about 9 months to get it to that place. Yes. How do you intend to deal with constraints which are which come with AI models, let's say, like you said you you want to train or you want to you want your models to know how the user acts and to uh to How do you say?

Um learn from him? So, usually models use the compact contexts. Yes. They got big. So, how do you intend to deal with such constraints? So, if I understand your question, I I think you're you're saying like uh I I think this is like another variation of like AI hallucinations. Is that Am I understanding you? No, it's actually Um contexts usually have a a limited size. Oh, I see. I see. Yeah.

They cannot The answers derived from from a compacted context are not usually not as precise anymore as from the original context. Do you have any back plans or mechanisms to compensate such I mean, a user would probably notice once in a while that his answers the quality of answers is not Yeah. So, how would you deal with that? >> Yeah, so um so so today I think when we look at on-chain actions, the amount of context that's used is significantly less than like if you're using a product like Claude or like like a a general broad agent.

Um we have not uh had a lot of problems of like uh filled filled context today. Um but our our goal is to to add uh to to rely on very small models and to um to to rely on a multi multi um multi-agent architecture. Um we are iterating quickly, so many problems to solve. Any other questions?

Yes. Are you fine-tuning the model or are you building it from scratch? We we fine-tune uh we fine-tune open-source models. Uh so uh I I think the the level of effort and requirement for for creating uh foundational models is is beyond the capabilities of startups in our space today. Um but also it's not really needed in my um in my perspective. Like fine-tuning models is fantastic and yeah. Thank you.

Yes. >> One more question. Kind of it's like the flip side of the risk. Uh are you thinking about doing something I call it semantic firewalls where you you would be sensing what the user is is typing and prevent them or ask them again, are you sure you want to do this because you know that this contract is maybe risky or something like that. >> Yes. See that that I mean that's one of the main use cases. Also, I mean we we want to get to a place where the agent also is explaining risks around specific contracts and can tell things like this is a um you know everything from this is a brand new protocol to uh this this um you know

this contract has a reentrancy exploit to uh this contract um you know that this contract is uh at risk from some sort of like a rug behavior. It it um it does uh simulations of of behaviors and uh today it relies on on public APIs, so it can only do that in contracts that are verified, but uh we think we can actually get to a place uh where uh even that is not necessary and where we can do uh that kind of risk analysis on on any contract.

Sorry, say say say that again. Yeah, please. I'd love I'd love to hear. Thank you for the presentation. Now it's uh it's finished. Thank you so much everybody.