AI Agents Are the Future of DeFi

Transcript

Hi everyone, thanks for having me. My name is Vasili, also known as Defi Vas. I'm a founder of Velvet Capital and um I've been in crypto since 2016 trading, investing, going through the ICO craziness, DeFi summer was on the VC side as well and uh managing my own D5 portfolio and uh I realized that the space is becoming more and more fragmented and complicated. So right now we have thousands of different D5 protocols, hundreds of chains and millions of tokens launching every day.

Now with Pump Fun and other launchpads, you have 10 thousands of different tokens that you just cannot keep track of. And so you really got to be 247 on with a team of analysts to be able to win in the space right now. Now, this is where AI comes to the rescue. And we've seen AI having a big impact on the space since um last year. Um obviously as with a lot of other things with crypto it started with a hype. Um so if you remember truth terminal and goat the first AI agent uh which raised funding from Mark Andre uh AI6Z which was the first agentic DAO and

then we saw an explosion of different launchpads started by virtuals and obviously at the peak we have tens of launchpads thousands of agents launching every day. So we've already gone through the first cycle of boom and bust and right now we see that there's a flight to quality um in this whole agentic space. You know what started as just Twitter slop bots now converging to real agents with real utility on chain. Now when we look at all the different DFAI use cases I like to think of it in two sort of axis here. First functionality and second autonomy.

So we have low functionality and low autonomy agents which are acting more of a research uh terminal agents and uh it's pretty simple to you know make sure that you have the right data um and the right models to be able to give research. Now what um happened last year and what actually brought this whole agentic meta to the spotlight was the high autonomy and long functionality agents. So we saw tens of uh bots on Twitter posting and u having some sort of a personality to them right it's uh the first try of um adding more autonomy to the agents without necessarily um getting any functionality

now where we um started to play at velvet and where we see the first use cases emerge is the co-pilot uh section so co-pilot section means that you have high level of functionality agents can execute DeFi tasks for you, trades and actually help you do useful actions on chain but at the same time they are not really autonomous. So you still got to tell them what to do. You got to prompt them and then only they will they will do what you say. Now the real question is when are we getting to high functionality and high autonomy agents?

So this is the real holy grail of DeFi where you know instead of having a team of analysts working for you 247 you have agents being able to execute and manage your portfolio chain. Now we were one of the first agents um out there. We launched uh and incubated Velvet Unicorn which was one of the first agents on virtuals. It was launched in October 24 before you know virtuals completely uh exploded with usability and um you know actually numbers of agents went to completely insane numbers um and what we realized

is that it's actually quite hard to uh to build this high autonomous and high functioning agents. So our journey started also with uh creating a Twitter bot uh making sure it has access to information making sure it can share different crypton news etc and then gradually we morphed it into copilot and right now we have a more complicated system. So this is what we started with. It's a very simple swarm um where we had a manager agent which takes a user input whether it's a direct prompt for you or if a it's a Twitter reply or a tweet tagging the agent and then it triages the prompt and passes to one of the

three workers executor analyst and newsmaker depending on the task. So it was very simple setup but it kind of proved that the multimodel and multi- aent model works um fairly well and uh where we got to right now is essentially this complex uh self-improving multi- aent organization and uh the way it works is we have a several sets of front ends one is our app itself where users can interact with agents and then it has a CEO. We promoted the manager. So now it's the CEO. This agent uh has multiple types of um categories in mind. And then

depending on the question, depending on the prompt, it categorizes the um task, passes to one of the agents, and then most importantly, it has DeFi specialists executing any onchain transactions for you and then feeding the feedback to the CEO to make sure that the main agent can check the result of the transaction and then improve the prompt further for future uh for future actions. So this is something that we built out um throughout several months and which is functional right now. So I'm very excited that you know we able to cover the whole end to end journey from

discovery to analysis to the actual onchain execution. Now let's get into more details on how this whole thing works. So this is the main CEO agent. Um, I put a few examples here of uh what it can do, but essentially it's a it's a separate model that takes your input and it has categories such as token specific questions, general crypto questions, non-crypto blockchain related question, token trade request, technical analysis, smart wallet activity, questions about virtuals because we realized a lot of users wanted to to uh to dive deeper there and uh questions related to our own product and platform.

And because it's a separate agent, it doesn't get overloaded by the context uh window. So the main purpose is just taking the question and funneling it to the relevant agent. What it allows us to do, it allows us to scale this to tanzo maybe up to hundreds of different types of uh queries and tests. Now another important thing is that LLMs are generally quite good at uh you know understanding your prompt processing the information and uh giving you you know general answers. They're not really good in managing D5 strategies or quant strategies. And so what we did, we uh included a quant statistical model um as

one of the agents uh which is trained based on 100 different technical parameters, back tested and then wrapped into an agent that whenever you ask um the main co about any sort of quant parameter, price prediction or anything related to um actual quantitative data, it funnels it to this uh prediction agent. Now the way it works is uh we have combination of uh neural networks cat boost models and uh essentially we we had a deep level of optimization to find correct predictions. So we actually um probably there's like a year uh or even more of deep research that uh led to you

know creation of this model. Um what is interesting is that it's very similar to the models used in traditional finance hedge funds. And so now we probably one of the first teams out there that's combining the you know LLMs and uh sort of um more of an entertainment uh agents with the actual serious um quantitative models here. Now another important agent is the execution agent because in the end of the day we don't want it to just talk we want it to do useful actions on chain and um essentially what it does it handles token swaps defy actions

bridging any sort of onchain transaction that you want to perform and the stack is pretty simple uh we have large language model that actually is extracts the intent now using this intent and we generate a payload um which the which we pass to our intent execution engine. So if some of you heard the previous presentation um about intents we do think that it's a very important aspect because in the end of the day LLMs are pretty good at identifying what you want to do. They're not as good as you know actually creating an onchain transaction for you. So using intense in the back end helps us simplify and reduce the

amount of hallucinations. Now just to bring everything together this is the stack uh that we have right now. So on the application layer our own DFA app is essentially acting as a front end for the users. uh we have several different front ends that we support as well whether it's other AI agents, asset managers, trading terminals etc. And then at the heart of it we have intent orchestration layer. So this is exactly the element I mentioned that helps you convert the user prompt what the user wants to do into actual onchain uh transaction down the line. And then we have several solvers that we're working

with including our own to make sure that the users get the best execution. Now this is one of the examples that I wanted to show you um which just recently we sort of put everything together starting from discovery. So we ask what's happening in the market today. Now the agent CEO takes this question, categorizes it, passes it to the researcher and then gives you a quick summary of um what's happening here and then even gives you like a few suggestions of interesting projects to to look at. Now we can drill down say okay what uh what tokens are looking bullish right or what tokens would you suggest um looking at? Gives you a quick

quant information as well. you can u kind of drill it uh drill down a little bit more to get to get a little bit more details on the project. technical analysis essentially it's a separate model that does a TA specifically based on the time series of um data both prices volumes and holders and then finally after getting all this information uh you can ask it to execute the transaction for you and so now what's happening the user is asking to buy this token with a certain amount of USDC

We take this intent and then we generate a payload sending it to the wallet. Now, right now, obviously, if you if you're a fan of external uh wallets, you can just use it uh right here in the app, but we're also integrating it into um um inapp wallets and vaults. So, that's uh just a quick example of what it can do. And then um this is um sort of the next step uh that we see right now in the whole agentic economy. Um I don't know if any of you are familiar with the virtual ecosystem but one interesting work that we're doing is creating this ACP agent commerce protocol to allow different

agent different agentic systems communicate with each other actually acting like a nation state of uh different agentic participants. So uh here every agent can request a service uh from every other agent and essentially communicate and collaborate to perform some complex uh workflow together. So for example, there's an agent that's running um a hedge fund. It can ask our agent for research and initial data. Can ask it for price prediction. It can ask another agent for um risk analysis. If there is an agent specializing on that uh and essentially combining inputs from

multiple agents, it can can basically go through pretty complex workflow ending with onchain execution. Now this is a quick example of um one of the tests of the quant model that we've run. Um so again something that we helping um is the um AI treasury management and um these agentic model was able to beat hod onlyly strategy um which is pretty cool and u this team this team appreciate it.

Now uh a few main takeaways here. I think those are probably three most important things if I if I were to basically squeeze everything um into those. First is the multi- aent setup um that we built and that we just covered helps avoid context overload and hallucinations. So we think of it as one of the most important ways how to build aic systems. It basically means that instead of having one large model and try to squeeze everything into it, we have a bunch of small specialized model and it allows us to make sure that we don't overload them and uh they actually each of them is conducting the task at

hand. Um second LLMs can be complemented by statistical models and quantus. I think this is also a very important point because I see a lot of people just try to use LLMs 100% different types of them without actually complementing it with um with quantus and um I spoke with a lot of people in defi and defi strategists we still think that there is a value of having quant models there especially since most of the parameters are deterministic and then finally the intentbased architecture can simplify a D5 actions So um this is quite important because in the end of the day we want useful agents who do things for us onchain and uh

using intent architecture is essentially making sure they have all the necessary tools to do. This is what I wanted to to leave you with. Uh happy to answer any questions. I think we still have a few minutes left and uh yeah happy to have a quick discussion with uh with you guys. Thank you. Thank you so much for having me. Just a quick question. Super impressive, by the way. This is really, really cool. Um, is are there active strategies that will trade like on your platform like without your input? Like you don't need to sign your transaction or anything like that? This is actually a great question. Uh, yeah, I should have covered it myself in the presentation, but yeah, just uh just thinking about

this um this sort of autonomy and functionality framework, right? uh where we want to get is to the top right corner. Unfortunately, I see a lot of people who promise you uh that there is an agent which you can just tell okay make me money on chain and it does it for you. I mean if it is true then it means uh everyone would use it and uh the margins would collapse to zero and uh I I I I think I think the um the truth is somewhere in the middle uh where you would still prompt the strategy different parameters and constraints that you wanted to run. So, for example, if it's a stable coin strategy, uh you can talk about your target APIs, your target risk exposure,

your target um I don't know, maybe even protocols that you want to uh to to choose from and then um and then run it, right? And then my prompt might be different from yours. So, in the end of the day, we're going to have different strategies, but they would run automatically and try to optimize for whatever inputs we we give them. Yeah, if you don't mind. So, what I'm thinking of something like even a stop loss. So you set your position but you want to have a stop loss but you're asleep and it triggers like what do you like are there any solutions for that? Yeah. Yeah. Yeah. I mean that's that's actually one of the use case. So what we have is uh we built a vault infrastructure. Um so the agent would be essentially managing a vault for you right you would uh you would define

vault parameters any constraints and then initial prompt for the agent. Um so the the future might look like you tell him okay um buy me every new token once it crosses 100k uh and thousand holders uh sell 50% when it goes to X or you know stop loss 20% below entry price and I would just run strategy. I think again just going back to my uh to my last point is a lot of people try to do it with LLMs whereas some of the parameters are highly deterministic you actually don't need don't need LLM or agents so I think it's very important that we actually combine LLMs with quant models it's for use cases like

So you say that like this challenge of uh an agent that if there was an agent that you could just start and run make money that would be ideal but it's kind of like this idealized version or view. And then the other end of the spectrum you have like these very prescriptive uh and necessarily iterative from the attention of the user um uh descriptions of what you want as an end user. Is there any room in between? Like is there is is it possible instead of at every step being like okay well it's done this now I want this it's done this now I want this. Um c can we jump a couple steps or or do you think you know because with with efficiency I guess um comes information risk in the sense that

as the end user um it erodess the longer that you may be away from it. So in your perspective it's really a subjective question. Do you think that how how necessary is the end user in this? because I worry about a future whereby we're abstracting the difficulty of DeFi only to become a slave to this terminal that like requires our attention to function. Are we just making a problem all again or is it possible to avoid? Yeah, I mean to your point I do think it's a spectrum, right? You you have highly prescriptive co-pilots where you need to actually define every transaction like the example I showed you. Um and then on the other extreme

just okay make me money I don't care how I do think there's a whole spectrum in between and it it's very important I mean prompt engineering is going to be very important because it will depend on the user how you set the initial parameters. uh but at the end of the day I mean what a lot of people do right now they go to chat GPT for example and they ask okay create me a prompt that is optimizing for X right so as a user maybe I I don't necessarily need to think of all the parameters and define all of them in advance I can still use the help of the agent to even define it initially then I tweak it and then I let it run um so that's probably where I would see it going yeah but I I do think right now we're still at the stage H

when the initial prompt will need to be given by the user and it's very important because that's the only thing that will differentiate your money-making strategy from my money-making strategy. All right. Thank you so much. Appreciate it. [Applause]