Panel: DeFi as an Agent Playground

Transcript

Today we're going to be having Ken from Uniso Foundation, Lucas from Mimic, uh, Gautier from Zifi, Stefano from Virtuals, all moderated by V from Dedalus talking on this topic. So let's please give them uh big applause and start the D5 as an engine playground panel. Hello everybody.

So we're going to have a panel now to see how we can use this AI agents, you know, to do fancy stuff to do DeFi and uh you know how how they can control our money, right? So that's kind of an interesting uh topic. We have uh two type of uh people's here like one building defi primitives, defi protocols, other utilizing u uh AI to to use these tools. So um yeah, let's uh get some intros first and uh see what who we have on the panel. Ken.

>> Yeah. Hi, I'm I'm Ken. I uh I lead applied research at the Uniswap Foundation. Um >> yeah, should I go more in depth? We >> Yeah, I think everybody knows UniS swap. So yeah uh >> uh hi uh my name is Lucas. I lead go to market at mimic protocol. Uh we're building an onchain programmability and automation layer where uh developers and a agents can create uh put complex logic uh onchain and abstract uh the the hard parts about um getting it executed and monitored.

>> Hey everyone, uh I'm Goier, founder at Zyi. Uh we're building a yield agent which operates directly from a smart wallet and enables you to get uh the best yields across different chains and and protocol automatically. Pleasure to be here. Hi, I'm Stephano. I'm the head of the US at Virtual Protocol. Uh at Virtual Protocol, we're building a society of AI agents. Uh we have a uh tokenization platform for capital information and we've also built uh an agent to agent commerce protocol. And so yeah, thanks for having me.

Okay. So, uh to kick off the session, um let's start with uh you know a more broad question. Uh what are the synergies between AI agents in DeFi that are making you excited the most? Ken? >> Yeah. Um [clears throat] honestly, I think you know if you think about where DeFi sits in um sort of like block space, right? you you think a lot about where uh you know the biggest gasers are are really trading it's um liquidations it's it's it's it sits so much in how much people use Ethereum today so when I think about agents I think about automating all of

these things I think about um more sophisticated strategies I mean obviously I'm I'm biased for unis swap purposes right um but where where people have really thought a lot about you know reapplication We've thought about like sophisticated um liquidity management strategies, right? It is, you know, the people who have the biggest edge um are not individual LPs. They're not retail LPs. They're not just like the average Joe. They're not, it's not me. Like I don't know how to do this and I'm just, you know, manually going through and like adjusting my position. Um but so many other teams, market makers, funds, they're doing this all on the back end.

And I think agents really kind of equalize that game, right? You know, once we start introducing more algorithms into agents, uh managing liquidity becomes a different ballgame. Uh and all the things we talk about for you know, minimizing um MEV, uh minimizing LVR, uh recapturing that it it becomes um yeah, it becomes a different playing field because agents really kind of level out how we engage with liquidity. So that's I think for me like kind of the most exciting thing, right? like how do we actually play a different game with the market makers?

How do we as like the average person, you know, compete with people with, you know, millions, hundreds of millions of dollars? >> Yeah, >> I think uh that's a that's a very good point. You know, I think one of the one of the greatest offers of DeFi is democratization of uh financial services and basically allowing everyone to have access to this thing. And now maybe we're bringing it like on the next level. So AI can help an average Joe to execute something more complex that otherwise he won't be able to. But uh yeah, what are your thoughts?

>> Um so I'm I'm very aligned with the topic of our panel where like DeFi is a playground for AI agents. So DeFi [clears throat] essentially is programmable money and uh you know agents if we want them to be part of the economy we need to give them access to this programmable money and uh that's the way how they can embed into you know our day-to-day economy uh what we're doing um and in terms of like DeFi specific when we're like talking about you know yield positions and and maximizing you know efficiency um I think yes these will give us like additional upgrade it's going to like additional layer of efficiency that we can achieve with those agents. Um and yeah like you know it's hard to monitor

everything 24/7. It's hard to react to some market events usually. Um I've heard an anecdote that you know a team uh a fund is hiring someone in Asia just to be able to um act on you know different market events when people in other parts of the world are going to be in bed. So I think a agents gives us this uh it's going to be a tool that will be used to to um enable more efficiency into how we operate in in DeFi. >> Yeah, this is uh basically operational efficiency. So I think the greatest example there is you know if you look at Tetar as the company with the biggest revenue per employee in history of humanity basically. So that's the kind

of the kind of going in that direction. Okay cool. Um Gotier. >> Yes. So I think AI is is interesting for at the moment giving suggestions to do specific specific things in DeFi but we've we've actually experienced trying building an LLM for generating yields and actually it was not having the results that we were expecting. Uh so then uh we built something which is completely rule-based um just like a basic uh agents and in script um improving the yield of users. So I think it's still uh it's still very early. Um but now with agents we're able to like abstract uh you know UX complexity uh automate a lot of different task uh

monitor the markets. So I think it's it's still like nent of what we're able to do and and as maybe as AI will improve maybe we'll be able to to use different types of models to do different things. But I I still believe like it's uh way too early to to implement AI on uh in DeFi in general. >> Do you see in your approach with like writing rules? So it's the traditional way of usually handling uh automation but do you see any improvements when utilizing you know LOMS and AI in this like old approach? Let's say it this way.

What what do you mean exactly? Like >> I I mean once you started using you know Oms and start building that do you see some clear improvements that could potentially like uh uh these tools can give you in in what you're doing? >> I think eventually it can give good suggestions maybe uh find data pretty quickly. Uh but for actually executing and and it's automated things I think it's it's just too risky for the time being. Um so we'll see how how things evolve. Yeah.

>> Yeah. I I kind of agree with with your take where like agents will serve as a again next level of um data aggregation and and analytics where it will provide you the information like you know the the analytics that you want to maybe use for your uh DeFi strategies and you know letting the agent like fully go and execute it is it is a risk if especially when we're talking uh about large amounts. So then like we at mimic have a similar like kind of rule-based approach where you have agents to execute the task but they're like very bounded in terms of like what they can do uh so so

they don't have like the full ownership of the execution but they are closing some boundaries uh of of you know and there is a human control over it. So I guess uh this uh maybe going to break the the dreams of some people that are going to be thinking that agents are going to be this completely independent entities that are you know doing fancy stuff but uh I don't know Stefano maybe you have some thoughts on >> I I'll I'll I'll give a different perspective. Um no I think I I actually want to build on something that that you and and Ken mentioned which was around um you know DeFi. I guess the word we use is democratizing access to financial services, right? And I think one of the reasons I personally believe, you know,

this has not been adopted at scale by by billions outside of of crypto circles is has a lot to do with UIUX, right? So you take something like Uniswap, which is super simple to use, very clean UI, um, but the whole end to-end workflow with wallets and all that is is not so accessible or so it's intimidating to an average person, right? And when you start to get into slightly more complex things like um vaults like like pers like you know all these kind of things I think I mean even for a lot of crypton native I I sometimes find like you know I find some new exciting vaults I go in I'm like I don't even understand how it works. I I some of them I don't even understand how how I deposit money. So I think I think at least today one of the

areas right for disruption with agents is is UIUX and basically truly enabling you know everybody outside of of crypto circles to on board to DeFi right one example I always like to give um is is a team or an agent called Momo it's a team behind Moonwell who who built kind of like an agentic interface for it and the beauty is if you go to their web app it's um it's like a fintech you know they have a a fluffy mascot, right? It's very clean and and and very sleek. Um, and it tells you that you just need to do three things. Open an account, deposit money, and watch your money grow. And there's virtually no mention of crypto or AI uh on on the platform, which I think is is is great. And I I

wish more builders in the AI/crypto space would focus on this. >> Yeah, definitely. That that that sounds great. And I mean everybody wants you know to have like this clean AI uh clean UI where you can ape your money in get like 10% out and be happy about it. So this kind of brings me to the topic of um risk curation you know because usually you abstract this away uh by you know uh giving the the risk curation part to a third party that's going to be figuring out okay here is how I'm going to manage this money. So if we go down the path of AI agents actually doing this curation game, aren't we going to

end up in the same situation like we had like a few weeks ago where the big question was how curators have actually skin in the game like how is the incentives not switched in a way where depositors are holding all the risk while the system underneath you know is just moving this money around and uh you know making some fees on top of it but bas but basically it's incentivized to seek the highest yield and maybe not so much to look at what are the underlying uh risk so how how to give agents skin in the game you have any ideas >> yeah I mean

>> hello hello yeah so it's yeah the rain is really loud this is kind of crazy >> yeah dude agents can't fix that um I mean to to Stephano's point right like there's >> like when we when we start thinking about like the the human interactions for agents in something like this it it becomes like um I don't know like how do you then write the rules for like risk programmed into models where you're like meant to make money you know like it it's it's hard to determine today to say like okay well you know there's only so much of like a risk threshold we can make for these these agents because we only wanted to no one's going to say like I only want I only want to make you

know this much money. I don't want to make more money. Everyone wants to make more money, right? Like it's it's kind of crazy to to think about like how we model risk in this manner. But I I don't know. I I feel like a lot of the algorithmic trading and the high frequency training that happens today is not is not all that um far off from what we're going to see into the future. I think we're going to kind continue to expand those models. We're going to expand the algorithms to be more aggressive. And I I honestly think like risk is going to going to like vastly increase. I we've seen this since like 2008 to today, right? Like we we go risk on until like there's a market correction and then we sort of like pull back and we're conservative, but then we like find new technology, we develop new

things, and like I mean we restructure products, but risk goes back up. You know, I I I feel like the parameterization in how we think about risk is I I mean it's just going to completely change the way we think about um like what is risk tolerance in the future of agents because it's not going to be individuals like uh training against each other. It's not going to be like people on the sales desk, you know, asking like, you know, what do we do now for for whatever equity. it's going to be like competitive advantage on geoloccation for um for those like slightest advantages and when when risk sort of like escalates that way the rewards are so much higher and I think the the way that we're going to see it happen is it's going to be this

acceleration towards um yeah towards hyper capitalistic uh um you know movements for agents like it's that's we we start we've already started to see a lot of this today uh and I I don't see a reason why that would stop. I mean it's it's interesting because you know right now when the the the money are lost you blame you you you have somebody to blame right in a AI environment you know it's a kind of it becomes a bit abstract like I I think even with self-driving cars you have the same issue where you know if an accident happens like who is viable for it right so wash I mean you you mentioned that uh you know you you put a lot of guard rails around this stuff so how how do

you see this problem. >> So [clears throat] I don't think that uh risk curation and like vault curation will go into agents anytime soon. And one of the reasons is exactly what we just said. There is you know a legal responsibility for funds that are being managed. Like you know it all looks nice in the front end that you know someone manages a vault of funds but we're you know often talking millions of dollars. So there's you know always some legal agreements there's some uh liability responsibility into it. So no one's going to be able like yeah if agent lo loses $10 million of of someone's funds like who's responsible for that.

[snorts] So I I don't think that the risk curation will switch into the AA agents. However, like as I said previously is AI agents will increase the efficiency and they will um reduce the the amount of you know time the curators need to spend on uh executing some transactions like this is this is how we see also like mimic playing into part of it where we it's a stack that increases the efficiency and can increase the the yield back to the users because those curators will not have to spend so much time. they're going to be able to automate some things uh and you know both agents and and um the infrastructure layer that we're building is kind of tend to improve the

efficiency but not to remove the curators itself. I I still think you know we're going to have human operators uh that will oversee um the the risk curation and the management of those vaults but you know these are just the tools that that will make this you know go forward and and make more efficient. This uh reminds me a little bit like when it comes to coding agents like with coding agents I think a lot of people think that uh you know you give the AI uh an idea like write me an app for whatever and it's going to do it but actually you need somebody to oversee it right so it kind of reminds me of that.

Um, >> can I can I push back >> because I I I think that's I think that's that's not that's not too far off. But like to like look at virtuals though, like look at how like there there's actual market makers and like ma like managed funds that are actively deployed on virtuals like you know following the mimemetic rise of like um you know these meme coins and I guess like that is >> yeah but we're talking we're talking memecoins not institutional capital right? >> No institutional capital. institutional capital like ask >> there's institutional capital in virtuals right now. We we do have and and I mean it's not the the the it's a it's a you know it's not the highest proportion right I mean a lot of it's still retail but uh but you do have and and actually I I think to your point we

we um uh we started to publish what we call agentic GDP so the amount of money in multi- aent systems um that are using ACP uh this started out 6 weeks ago it's now reached about $130 million so I mean by by defi protocol standards is still pretty low, but like it's it's it's getting there and it's growing about 10 million per day. Um, >> it's it's a decent amount. I mean, but it's still like an experimentation phase, I guess. And like Yeah. >> Yeah. I would agree with you. We actually um did a partnership with with virtuals for the agents uh to agent funds. Uh so so we we released it and now it's the total TVL there's only

about 500K. So it's uh it's still it's still very nent and and the way it's built like there's still like some some offchain components. So it's uh no it's it's it's it's uh eventually shaping up. Um but but yeah to to >> jump back on on on the initial question uh for example uh on our side the way that we add opportunities and different pools and vaults uh is all done manually, right? We have we have two quants that that analyze the market and tell us oh there's this opportunity and this one you should you should we should take off. Uh but we thought about actually like using AI to just analyze the market um and like maybe create reports for us automatically analyze Twitter maybe like for example like the

FUD that just happened with XUSD have you know like different types of agents that would monitor this and give us alerting system and then we can take action. So there uh the AI itself would would just help you know us and curators uh you know take better decisions and maybe take this decision faster. Um yeah. >> Did did it work during the XUSD JPEG like the the the meltdown? >> No, we saw it and yeah we rebound installed the users funds uh manually. >> Uhhuh. Okay. But you managed to to see it on time.

>> Yes. Yeah. Yeah. >> Amazing. That's great. Okay. Uh Stefano, you you have some thoughts on top. >> Yeah. So, I'll add something. I mean, going back to the original question around um you know, trust and guard rails and all these kind of things. I think we um about a year ago, you know, uh maybe maybe 11 months ago with this whole hype around um agents that could grow your wealth while you slept, right? You'd be like, "Hey, I I want to be I want to retire by tomorrow morning. Click the button and and you know, and and do so, right?" I think we quickly realized that that was not going to happen and that distinct agents or or distinct pieces of code were not going to be able to do that. And as a matter of fact, agents would have to would

eventually become highly specialized at doing one particular part of the workflow really well. Um, and and that's what led us to build ACP. And I think part of the philosophy behind ACP and and and we're not there yet by any means. think it'll take some time to get there is that um you know we we've kind of like instigated or or incubated I don't know maybe it's the right word you know some of these autonomous ideas like hey what if we had an autonomous asset manager or an autonomous fund of funds or whatnot right and these systems would have to be multi- aent right so for example you know you would have to have a group of agents that are looking for alpha and some of them are looking for alpha you know uh for for kind of lowrisk investments and others are

looking for medium risk and others are doing the super DGEN, you know, 10x leverage per trading for memecoins, right? Um, then you have a bunch of agents that are looking at risk and and some of them uh are looking at like platform risk or contract risk, you know, rug risk. Um, actually the the guys at Cambrian are building one that's looking at financial risk, especially for the more kind of like risky bets like if you're if you're doing PERBs, you know, you know, is a liquidation risk acceptable, right? And then you have other agents that are actually focused on execution. So, you know, the ones that are um let's say very good at at, you know, uh uh rebalancing your vaults, right? Or the ones that are the best s just finding, you know, the deepest liquidity, the lowest slippage for swaps, um you know, etc., etc. And

um right now it's very very uh ironically human orchestrated like we basically pick the teams we want to work with and um and and you know, we we kind of seed the ideas and we try to make it work. In the future, we expect that um you're going to have obviously more than one agent that can do one thing and effectively they'll end up having to compete for you know for doing the task whether it's on you know I don't know cost or quality or or whatever and and then this is where trust course will be very important and one of the things we're integrating is 8004 as a way to to to have that kind of verifiable trust of these agents that you know over time the orchestrators of these multi- aent systems will be able to say Hey, here's

a user prompt. I need, you know, five agents that do these things and I know which ones to pick based on, you know, the risk profile and and the the price sensitivity and all that. And then the trust score of of these agents. >> Okay. So, you still have the the human kind of decision, the final decision on top basically that uh you know, somebody needs to make and that is probably going to be the one responsible if things go wrong, right? Um we are kind of running out of time so I I wanted to to dive uh a bit deeper on the topic of when these agents make decisions they base these decisions on data right so now the question is like and this is a kind of a trade-off between uh you know the

verifiability and uh openness and all this but if you know what data these agents are basing it on isn't it possible to manipulate this how do you actually make sure that somebody's not going to attack you know the the system through you know basically it's an Oracle attack of some sort but how do you make sure that the agents you know made the took the right data in order to make the decisions like how do you protect there like Lucash you have some ideas >> yeah so it's a very good question uh so obviously I don't know specifically how agents take the data like how they analyze it and how they can manipulate it but the approach that we've taken at mimic is that the protocol itself when you create a task is monitoring the data

like the onchain data, offchain data sets, whatever you want to plug in, whatever you want to follow. So then you have like a verifiable uh data that whether it's a user or a agent that's uh triggering the task uh that you know it's it's based on the verify verified data. I guess like yeah I'm I'm I'm not sure how to if an agent has got access to the private keys if you know they're basing their execution on some sort of uh data if that's manipulated then obviously you need to have some safeguards in place right without without this you know you're you're very exposed and your agent can get hacked.

Uh so so it's super important to have this execution layer secured programmably on chain so that you you limit those boundaries and that the the data is also verified on chain uh before the execution happens. >> Mhm. Got it. Uh Ger you have some ideas. >> Yeah. Yeah definitely. So um for now we've been running six six months we've been improving our model model and algorithm um like almost every week and you know we're always asking ourselves like is this something we should uh we should completely publish and and but eventually you know like you don't know like if you're going to rebalance there's going to be like someone's going

to frontr run you or any type of other other aspects but what we're doing for right now is u we're going to implement ERC we actually already implemented ERC C84. So for each rebalancing uh check, we're going to create a ZK proof, upload that on the validation registry. So that will just prove that the rebalancing that we've done based on certain conditions has has been done uh correctly. So but this is actually a great question because we're trying to figure out how we can provide you know the most transparency without also publicly like providing too too much risk for the for the end user eventually.

>> Yeah. Yeah. It's it's a trade-off, right? So Ken, uh, you have some thoughts on the topic? >> Not, not really. I mean, I think like, um, [clears throat] I think there's a lot of really cool solutions with like Brevis coming up with, um, uh, a lot of the using ZK proofs to to kind of like um, uh, kind of index a bunch of data together. And I think that's really interesting. I think there's a lot of new models for um, onchain data, right? like taking T- Wops and VWOPS directly calculated from onchain data. Um, so that you don't have to rely on like a centralized oracle. Um, but yeah, I mean I I think generally like this is this is like a field that I'm I'm so interested to see like what's going to come up. I I

have no I have no strong thoughts actually. It's it's going to be cool. >> I think Stefano probably going to have the most to say on the topic, right? Um, I'll I'll preface this by saying I literally have no bags or skin in the game, but but you know, Cambrian and and and actually I I I mean I know Sam pretty well, the founder, and um we're working closely with them, and I think we were very excited about what they're building, which is basically this this verifiable data network for for agents, right? Um but I think I think my view and and the way that we're approaching this at virtuals I think in the short term is that um you know we're not yet at this stage where as I was saying earlier we can tell the agent or system of agents hey I'm done working I want to

retire by tomorrow make me rich overnight um I think there's still a lot of um there's a lot of thing I mean data is one problem we still have hallucination issues issues with agents right and I think one of the things of how we're thinking about in the medium medium to maybe short term of fixing it is that you still need to have some level of human in the loop whether it's a curator or whether it's the actual user um just confirming certain things, right? Uh because a lot of times whether it's a data issue or a hallucination issue um today there's still a high risk that um the agent's going to do the wrong thing. We've had some examples in our community that, you know, the prompt was not maybe crystal clear and the agent went and did something that the user didn't really want it to do. And so

the solution today is just say like well you need to maybe double confirm that that's what you want to happen. And I think I don't think it's necessarily a bad thing because I still think that there's a lot of value in um you know bringing AI and using agents to do things that um agents can do better like agents can consume and process data and and kind of come up with decisions much faster than or or at a greater scale than humans can but you still may need today to to have you know a human in the loop to just make sure everything's fine. Yeah. Um I think we are we're running out of time but uh yeah uh if you if I need to summarize basically like I can uh say like even uh you know after this topic I have like two uh two

people fighting in me you know um one is saying like oh my god like this is never going to work like it's just so complicated but then on the other side I see what is happening with uh AI coding agents like when it comes to like writing code and if somebody told me like six months ago that you we want to have these agents writing like uh thousands of like hundreds of thousands of lines of code and shipping software. I'll be you know I'll never I'll never believe that. But right now it seems that things are happening and you know engineer software engineering is never going to be the same after the AI coding agents and I think uh you know from from this conversation I feel like we are heading with financial uh uh uh

infrastructure we're heading this way basically agent allowing us to be more effective uh making more effective solutions still humans going to be involved like I don't think it's going to be like completely on autopilot but then you know the uh the bar for getting into the game is going to be so much lower. So uh yeah, thank you very much for this discussion. I hope that uh that was also valuable for you and uh see you in the hallway. >> Thank you. Thank you.