Trusting Agents With Capital: The Gearbox Playbook

Transcript

[applause] >> Thanks for coming. Pretty cool to see you here. And let's start from one number just to go through this presentation. So, this is zero. And today I want to say why it's so crucial to talk about. The first part, we are in Gearbox like a maniac on security. So, we have a zero security incidents last 5 years, which is cool. The second part, we have zero bad debt for this part. And of course, when we talk about how to manage capital with agents, we want to bring all robustness, safety, and

security. Because as you know, agent could behave in very inappropriate way. And zero could be how much money left on your account if you do not secure agents properly. And my talk today is about my personal experience, what I learned piloting different experience different experiences in Gearbox last 2 years. We start to cover this topic in 2024, building some prototypes on EigenLayer, and learn how we do it properly. In this talk, I want to also talk about Gearbox agentic framework and how we perceive is the best way to organize capital management. And at the end, I can give a

really proper way how you can set up right now today a self-custody AI wall. So, it's curator-less autonomous way to manage your capital as a structured product. And at first, I want to talk about numbers because we should all together understand what we are facing. Regarding to BCG, there are 19 trillion will be moved and tokenized on chain in by 2033. It's a huge amount not only not only as volume. Just imagine how many different kinds of tokens we will have in the future. And the second part is more

crucial. Today 62% of entities are experimenting with AI regarding to McKinsey, but only 10% use them in production. It's a huge huge gap. And my question to you, why? Why we are so in experimental stage? Why we are not so moving forward to real production task? And to answer more, I believe and talk about the vision how DeFi should be done. Three weeks ago, maybe you have seen in a Twitter it was a post that AI is so cool by finding different vulnerabilities that DeFi is not the safe place anymore.

Previously we faced with many hacks in industry, but right now people are pretty scared. And as you can see, Manuel Arros, who is a founder of Open Zeppelin, pretty well-known person in industry regarding to security, wrote this tweet, "Hey, DeFi unsafe." And I think, of course, we are in industry, but imagine what happens when people who manage real capital could check that. How we can really solve it and solve more complicated stuff?

And I have a vision that if we heading to DeFi 3.0, we have two theses. The first one that in the future we will switch from complicated smart contract system to system which will use Ethereum as a settlement layer. Previously, for example, if you consider some pretty complicated uh smart contract system, there are a lot of things inside like internal balance etc. This less analytical logic. When we go to P2P, P2P, we just settle only one smart contract which supports a credit line and only two person get an access, lender and borrower. It's a

pretty simple to do through the formal verification. And after that, you have a proof that your smart contract has a limited amount of users. The second part which is more interesting that there are so many assets, it's pretty hard to set up a market pool market for them. RWAs, uh things regarding to factoring etc. etc. It's a very complicated way to build a market on top of them. So, and the third part that RWA is not good design to deposit and redemption. Uh if if there are some delay, but it's another topic. The second point, if you have a settlement, if you have one-to-one contract, it's very significant for both parts to make due

diligence and autonomous management. Otherwise, we will end up with millions of assets and no one can really check what's going on. Right now, we use a risk curators, but I think the future is curator-less when anyone can use their agents to check and make due diligence regarding them. And these agents could make a deep research, they could really fast response to market changes, and it's what we are talking about. And why autonomy is matter? So, I think many of you use agents to make a research, but if you make an autonomy, you have a much better speed to really follow the market. You can really scale

them up. You can really analyze multiple stuff and today I will show you a real use case what we built. And the second the third part it never sleeps. So, I think right now you can see people who check in Twitter for decision-making. Agents works autonomously. And in Gearbox we go through the whole complex problem and build a Gearbox agentic loop with six six stages. So, we start from discovering when we make a kind of initial analysis of different options, different assets, different targets. Then we start and deep into analyze to make it diligence and agents are pretty cool and I'll show that to go through the whole process taking

different pieces of information into one context regarding from Twitter, uh internet, etc., etc. But because LLMs are not fast, we need one more stage and it's a kind of new because LLMs takes time. So, we need to have a short list of potential candidates and then we go to propose. We try maybe to go into this position and it's not all time happen because this market is a very very fast changing. So, if you find it's profitable for you until the price is, I don't know, 1000, then the price is changed until you do this computation.

So, it back and forth. And then of course it's more crucial how to make execution safely and continue to monitoring. And this is a demo how our AI agent works. So, you can see this is in our agentic app. You can see my two positions. One in is USDC and one is a Midas LP and strategy. And we ask agent, "What is the best way to deploy 10K of capital?" So, I just write a normal request how to deploy 10K of USDC and you can see that agent continuously go through all data we collected. The biggest problem we face is there is not enough good source of proven data. So,

we built our own indexer and right now you see it collects different data, provide graphs, etc. Which is pretty cool information to analyze and make what I told during the discovery stage. When we have a sum of few candidates, the agent right now is going to web search, go through Twitter, but it's done so fast that even you can't go across the old town. But in this example, I think around 100 documents were analyzed and the agent finds that some our pools is not secure enough because there is some incidents with RLP token. Then I click just to deploy and here you can see a pretty cool way to

deposit. So, after all this stuff, I can make a really decision what to choose. And as you can see, I spent only 5 minutes, but the agent were able to analyze at least 100 documents related to the market. But it's not enough. So, this happens on Gearbox web servers and I'm a security maniac. So, we go next and we uh presenting an MCP server. So, you can run this loop into your cloud code, your codex, or any agent you like. And the logic how we want to achieve security that when you have a kind of decision, it makes a verification link. So, you

never sign what you do not understand. Human in the loop is crucial at the current stage and I believe it would be for a long time. So, you click on this link and you open a special website when you can understand written by human language what you are doing. If someone would be able to to some toxic context to your agent. If someone could fool it, you are the last mile to check that everything goes wrong. And this verification use a normal web 3 security stuff. So, we do not rely any server. So, you can deploy it into IPFS. It's immutable code. So, it's a pretty safe that you can check

what's your agent suggest you. But the most interesting part, I think, in our system is how we can switch and do the next step to autonomous capital management. As a logic here that in Gearbox, we have a battle tested what I started before for 5 years the model. We know how to compute total value of your credit account. And built on top of this idea, we created a bot APY. Bot APY is a way how you can make a on-chain harness and allow some operation with some risk budget. So, as you can see, this is a window from our application when you

have a position and you can allow your agent, which run in a cloud and bring different news sources, Twitter, etc. to manage your capital. And you set up a risk budget because all time when agents will try to execute transaction, we check that today the risk budget is enough. So, we compute total value before transaction. It's maybe 100,000 and after. And of course, you should give it because there are some slippage, so you can't really do the same, but you can really minimize potential losses.

And if you can minimize potential losses using the agent-driven management, it makes a pretty cool way for institutional players because it's a very easy to hedge. If your loss is kept maybe 2.0% you can simply buy insurance and it would not be some huge amount. And the third part, if you failed with your opsec, as you know last day it was a lot of incidents. For example, the key of your agent was leaked or something goes wrong. You still kept with this daily risk budget. You can recognize the problem and you never get the access to

whole amount of money. And another part which is pretty complicated when you build a agent who make decision through this news etc. is how you can make its memory. It's not enough just to read different Twitter sources because as you know, it's a very very manipulatable things. Previously people tried to manipulate price feeds to increase price etc. Right now you can write a lot of tweets to make a toxic impact to your agent. And solution we are working on that we have a graph of data and each source has a kind of credibility score to help the

system and understand is the statement have some kind of credibility or is totally incorrect. And in this case, maybe someone read the news when people add some comments in the quote and quote changed behavior. With these things it at least protection how to make agent to follow this way of thinking. And another part using graph memory, it's a pretty cool that it's a very extendable. So if you are institution, you have your own due diligence department, you have your own methodology, you have years of evaluating different assets. And this data is easily pluggable to this graph and help agent to make a better

decision-making. Previously, when we talk about capital management, we try to beat machine learning system, which should win in all cases and provides a high efficiency. But right now, I think we will change a little bit with LLM, we can provide some manifest what we want to achieve, which areas we believe will grow further, and then this system could follow your manifest and try to do the best. Why it's so uh different, so two different institutions set up the same system, but provide different goals could achieve them.

And the second part related to the graph, it's why agents is much much better than all existing risk curators and humans. Why? Because today assets are very complicated. This started in 2021, and I definitely remember the way when token symbols was only three letters. It's like or four, USDC. Right now, we have tokens like Woo, Moo, Boo, Base Wrapped etc. Complexity is coming. And to find all connections between tokens, some of them using collateral in Gearbox, then it connected with Morpho, with 3F, and multiple protocols. Imagine one asset

could be interconnected with 100 companies, and you can't easily to understand such as interactions. You can't easily make these things. If you build this graph properly on the depths, you can make different simulation, and you can teach your agent how to behave. So, you can't even find, and it's crucial, if some asset could be affected. You remember maybe this extreme finance hack when Elixir based on them and then some kind of risk creator ban of them etc. etc. It's like a domino effect. And of course it's a pretty cool to have such a depths and then you can set up a emergency exit. If some asset is failed, you can

immediately find the interconnection use this deep research and save your money. I think it's the best thing we are building on. And as I promised, so how you can make even better system if you are not interesting to use leverage, but you want to have a kind of better asset management for you personally. So the core idea is to build a structure product based on Gearbox stack to have um some harness regarding your management.

So for example, you have a pretty stable source of yield. Let's imagine you have a 10k uh USDC tokens. Then you split it into two categories. The 90% of this amount you put into the pool and because they are in a form of yeah, a USDC, they generate yield for you. It's your passive side. Then you set up a credit account which is managed by agent and could go through different positions investing the different tokens and opportunities.

This agent could borrow money, but now it's a quite straightforward that if agent would not really provide the same performance as aUSD tokens, it would be liquidated. We use the word liquidate as a lending protocol, but here it works as a stop loss. So literally, you have a 90% of the pocket saved, then your agents borrow from your pocket and try to outperform. And if this agent do not reach the performance goal, such position has as on-chain stop loss, which is also well-tested and battle-proven for last 5 years. And in

this case, you should not rely only on stochastic decisions by the agent. You have a pretty robust on-chain system which create a harness and really strict rules where you could not underperform. You should compete and overcompete our. And this way, it's a very good way how you can launch it tomorrow just for example, me personally want to do that because I have not enough time to search Twitter and everything. And then it's automatically finds the best way how to overcompete. As you know, the best way how Google teach the robots to play Go.

Robots play with themselves to find the better strategy and learn from each other. The same we can do in finance and this is one of the way how it should be done. So, this is all things we believe that we will launch in experimentally our Gearbox agentic. If you have Oh, I think I have already on time. If you have any questions, maybe we have half a minute. Thank you so much. Thanks for your attention.