OKX Ventures recently held an online sharing session (Twitter Space) with the theme of "Autonomous Agents Reshaping DeFi", which delved into one of the most exciting intersections in Web 3: the rise of DeFi autonomous agents.

This discussion went beyond the early conceptual frenzy of AI chatbots to address the core question: How can autonomous agents create real value, manage risk, and fundamentally reshape the user experience in decentralized finance? To gain frontline insights from the builders, we invited four industry pioneers who are shaping the future of Agentic Finance:

• Cambrian Network's CEO & Founder Sam
• Almanak's CEO & Founder Neo
• Giza's CEO & Founder Renç
• Makina's Product Lead Colin

AMA Summary:

1. AI isn't an incremental improvement to DeFi; it's a paradigm shift. Its goal is to transform DeFi from its current complex, product-centric model to a simple, user-centric, personalized service. Ultimately, users' financial goals can be achieved autonomously, without requiring in-depth technical knowledge.
2. AI's clear division of labor: "off-chain brain," not "on-chain hand." Currently, AI's role in DeFi is strictly limited. It primarily serves as the "off-chain brain," performing complex reasoning, data analysis, interpreting user intent, and generating deterministic and verifiable strategy code. AI itself does not directly touch or manage on-chain funds; its ultimate execution is based on auditable logic similar to that of traditional finance.
3. Safety First: Manage risk through a combination of "human oversight + technical guardrails." We must prioritize safety and risk control to address user concerns about uncontrolled AI. The core approach is that AI operations must operate within code-enforced "guardrails" preset by human risk managers, and the policy code generated must be fully auditable and verifiable by humans. This ensures that AI decisions are controllable and traceable.
4. Serving two types of clients: improving efficiency for institutions and lowering barriers to entry for retail investors. The product targets both institutional and retail users, but in different ways. Institutional clients such as hedge funds and DAOs use AI to significantly reduce the cost and time of strategy development and operations. For retail users, the goal is "radical abstraction"—hiding all the complexities of DeFi. Users simply express a simple financial goal (such as "I want stable returns"), and the agent handles the rest.
5. Ecosystem synergy: The application layer and infrastructure layer develop together. The realization of Agentic DeFi requires a complete ecosystem. This includes not only user-facing policy application layers like Giza and Almanak, but also the "rails/settlement" layer like Makina that provides a secure, cross-chain execution environment, and the infrastructure layer like Cambrian Network that provides reliable, verifiable data "fuel" for agents.
6. The ultimate goal: to democratize professional financial strategies. Through AI agents, we aim to break down the barriers that have limited access to complex quantitative strategies in traditional finance. Hedge fund-grade strategies, which previously required millions of dollars and months of development, will be made available to everyone at a fraction of the cost and speed, truly achieving financial inclusion.

Original AMA questions and discussion:

1. Introduction to products and main focus

• Sam (Cambrian Network): I began my career as a cryptographer at a US national laboratory, primarily reverse engineering cryptographic hardware. I then earned a PhD in reinforcement learning at the University of California, Santa Barbara. I then founded my first company, Semiotic Labs, where we were the core development team behind The Graph Protocol, focusing on AI, verifiability, and The Graph's payment system. During that time, we did a lot of work with agents. For example, in 2022, we released the first reinforcement learning agents for dynamic pricing within The Graph. In 2023, we released the first publicly available blockchain data terminal, allowing users to query real-time and historical data using natural language-generated SQL. In 2024, based on these experiences and our conviction that AI would have an immediate and significant impact and that cryptocurrencies would become increasingly important in the global economy, we decided to spin off Cambrian from Semiotic. Cambrian focuses on providing on-chain and off-chain financial intelligence. Providing this intelligence to agents is our beachhead market.
• Neo (Almanak): I've been in this space for nine years. Before founding Almanak, I ran a data science and consulting firm specializing in DeFi, trading, and crypto asset management, so I'm very familiar with how this space works. Regarding Almanak, we've been in the market for four years. We like to call ourselves a vibe coding company , but you can think of us as the Cursor of DeFi. Essentially, we use AI agents to discover and build complex trading and asset management strategies. These strategies are fully verifiable, deterministic code. You can think of them as the same strategies any hedge fund would use to trade.
• Renç (Giza): I have a background in product and marketing. Before founding Giza, I was Head of Product at Johnson & Johnson for five years. During that time, I built smart contract systems across various financial use cases. I'm fortunate to have a team with backgrounds in machine learning and data science, so we're approaching this more from the perspective of machine learning and AI, with financial experience. We've been building Giza since 2022. Giza builds agent applications for automated finance—autonomous systems that can execute complex financial strategies on behalf of users and institutions with zero operational overhead. I like to say it's our version of "banking the unbanked." We believe financial exclusion isn't just about having a secure account to store your ever-inflating fiat currency; it's about being isolated from opportunities. Are you able to adapt to changing markets, capitalize on these enormous opportunities, and mitigate risk when necessary? These are the questions we seek to answer. Our work at Giza is to democratize all of these capabilities.
• Colin (Makina): I'm responsible for product at Makina. I joined the team about four months ago. I've been in crypto for over a decade. My background is originally in traditional finance, but I got involved in DeFi around 2016 and have been building products ever since. At Makina, we're focused on institutionalizing what we call "DeFi execution." Beyond strategies and vaults, we're really interested in creating a secure and reliable way to interact with any DeFi protocol or any EVM. This is critical for anyone trying to run a strategy, whether it's a more traditional human operation, a more passive or automated strategy, or one built through an AI-driven agent approach. We approach this from multiple angles. First, we focus on the role we call "operators." This is similar to the "curators" you see in other protocols. They enable transactions to be executed securely while maintaining control over what they can and cannot do. Beyond that, we ourselves use AI extensively to improve the user experience, such as providing better recommendations, better understanding what users are doing, and researching different ways to integrate new protocols to ensure that whether humans, agents, or other types of algorithms are operating the vault, they can get started quickly and create maximum value in a secure manner.

2. What inspired you to start your current project? Why do you think AI will bring value to your product, and what is the main value proposition?

• Sam (Cambrian Network): I completed my PhD in December 2019. Reinforcement learning is incredibly hot right now, but in 2019 we were in a bear market for it. This is one of the reasons we focused on fully homomorphic encryption when we first started the company. When GPT came out in 2022, I was initially as shocked as everyone else. But I actually thought we were at the beginning of a bubble—and I know many people still think we are today. But by 2023, a year after GPT's release, I was seeing continued progress, and I developed a deep conviction, which I still hold to this day, that we are at the beginning of a new revolution. The previous revolution was the Internet Revolution. Before that, we had the Silicon Valley Revolution, and before that, the Industrial Revolution, and so on. So, we are in the early years of a new revolution that will not go away. I encourage everyone here to prepare: AI capabilities will double every year for the foreseeable future. This will impact everything, every aspect of our lives. It's already begun. Beyond this belief, I'm also involved in DeFi. Back in 2021, my previous company founded Odos.xyz, which we spun out. It was a DEX aggregator. So, I have a deep belief in financial applications and the financial freedom and literacy that cryptocurrencies bring. During the pilot projects and experiments mentioned in our introduction, one of the most challenging things I noticed was how difficult it was to access data and information about what was happening on-chain, as well as other relevant information crucial for both on-chain and off-chain financial decisions. This is crucial for financial decision-making. That's why we focused on Cambrian. We believe that every project engaged in agency finance or autonomous finance needs reliable, fast, comprehensive, and verifiable information to feed their agents. This is critical to the success of these projects, so we decided to focus on financial intelligence.
• Neo (Almanak) We like to call ourselves "AI for DeFi." Regarding our inspiration, Almanak began as a company using AI to optimize trading and asset management strategies. We've been working with large asset managers and allocators, so we've always had access to large amounts of capital. Almanak has been around for four years. Three years ago, when the ChatGPT craze began, we knew it was going to be really important. So we asked our large clients, "Hey, what would convince you to trust AI with your money?" They said—and these were particularly large allocators—that they would never deposit more than, say, $100. They were extremely afraid of AI manipulation, indirect cue injections, and all sorts of "unknown unknowns." Or, simply put, if their money went down, they wanted someone to sue. So, when talking to these institutions managing billions of dollars, we asked ourselves, "Okay, what is AI best at?" Today's AI is best at coding. It can code hundreds of times faster than an average person. AI is also very good at reasoning; it can process information trillions of times faster than a human. We took these two characteristics and applied them to Almanak. We've created an "Agentic Swarm"—or team of agents—whose objective function is to write a high-performing strategy, identify market opportunities, process market dynamics, optimize existing strategies, and communicate all this information back to users. In our ecosystem, AI collaborates with users. It provides you with ideas for strategies, optimizations, and ultimately, the code. However, if something goes wrong, you're the one hedge funds will call. What we've created is a method that reduces the time to market for developing any complex financial strategy from months to minutes. Furthermore, we've reduced the cost of developing such strategies from millions of dollars to just a few dollars, or even less than ten dollars, depending on the complexity. Once a strategy is created, it's exactly the same as any hedge fund would use. It's deterministic, verifiable, and can be backtested, simulated, and deployed—so you know what will happen. The AI never touches your funds. It simply enhances the strategy creation and discovery process, but never touches the funds. So far, this approach has been effective. We've seen the confidence of large allocators, and our total value locked (TVL) is currently $160 million. What's also very important is that once a deterministic Python strategy is created, you can wrap it in a vault. These vaults are fully composable—you can put them on Pendle, Curve, and so on. So that's pretty cool, too. We like to think we've created a new asset class called "tokenized AI vaults." Again, the AI never touches the funds, so large allocators feel comfortable depositing funds here. They know who to go to—they'll go to you, the vault operator—and you're just using Almanak as a tool that makes coding 100x faster and understanding a billion times faster. Also, as Sam mentioned, we're also focused on creating financial agents. Our agents are fine-tuned for quantitative reasoning, so we ensure they're as smart as, or even smarter than, any other quant in the industry. But our inspiration comes primarily from working closely with large allocators, not from trying to meet their needs on a practical level. We simply ask them, "Okay, guys, what do you need? Where would you invest your money?" and then we build it.
• Renç (Giza): As I mentioned, before founding Giza, my partners and I had been building smart contract systems across various financial use cases. One thing was clear: while these self-executing contracts unlocked the promise of open finance, we frankly believed that, in its current state, the pace of innovation was too slow to remain competitive with traditional finance. This was the primary driver behind our search for ways to bring complex off-chain computations on-chain to significantly enhance the capabilities of decentralized systems and the user experience of interacting with the decentralized finance world. Since 2022, we've been deeply researching verifiable AI. We were doing it before it became cool, explaining its importance, particularly in financial use cases. We explored all possible machine learning hotspots and financial use cases in decentralized finance. For us, the value of AI is twofold. On the one hand, generalized intent processing is key to understanding what users want to achieve financially without requiring technical input or autonomous, specialized actions. On the other hand, it's about executing complex adaptive strategies on-chain with precision and zero overhead. This second part leans more towards small, in-house developed machine learning models and traditional finance algorithms that are used repeatedly and are fully interpretable, verifiable and customizable.
• Colin (Makina): Our story at Makina begins with Dialectic. Dialectic was our design partner; we're now independent from them, but they had a realization while building their own fund. For those unfamiliar, Dialectic is a very active investor in this space and has been one of the earliest and most advanced players in on-chain yield strategies since 2021. They manage many different things through systems they've built over the years. One thing they quickly realized was that in order to compete in this space, to make money, to outperform, and to attract more depositors and limited partners (LPs) to those funds, they needed to outperform other strategies on a risk-adjusted basis. To do this, they built a lot of different tools that leveraged scripting. One of the technologies they used was an open source project called Oiler, which they actively contribute to. They realized that many of the tools they built would actually thrive better as open infrastructure. That's pretty much how the Makina story began. We essentially brought it to market, partnered with them, and are now expanding to other operators in this space. We want to support the future direction of development, which is moving towards more automation. This automation will be heavily dependent on what's happening within the blockchain, what's happening in the macro environment—wherever you get that information. And the best way to approach that information is what we've heard from everyone here. We're approaching this problem first from the perspective of DeFi and financial infrastructure issues, and then looking at where best execution, best decision-making, and best data analytics can be applied. And that's clearly autonomous agents. We recognize that, as we heard from Neo—which was a great opening statement, by the way, about some of the concerns people have about handing their money over to AI—we approach it in a slightly different way. But we firmly believe that as these technologies get better and people begin to understand, we can complement and expand our offerings while also addressing some of the major cost issues that occur within the asset management industry. So we're big believers in DeFi, big believers in Ethereum, and big believers in AI and its advancements in these industries.

3. Who are your main customers now? What are their pain points?

• Neo (Almanak): At Almanak, our product needs to solve two problems. We have to solve the problem of how to provide complex strategies and vaults. So, we've been working with many different DAOs and every curator on Morpho, like Stake DAO, MEV Capital, Block Analitica, Gauntlet—all of them. When it comes to DAOs, we're in talks with most of the top 20 DAOs on DeFi Llama. Why would they use our product? They'd essentially create vaults that leverage their assets. Let me use Ethena, one of the largest asset managers, as an example. Imagine you could have a USDe vault that continuously optimizes and seeks the highest USDe yield across all DeFi protocols. We're talking to those people. We're also talking to a lot of new projects. I don't know if you're paying attention, but there are a lot of complaints about high FDV token economics. So, at Almanak, we also allow projects to leverage AI to launch their own liquidity provision or trading strategies. Users can simply use our algorithms to launch a market or a trading competition. Last but not least, there are the average user, where the capital comes from. So, I just explained the supply side of the vaults. The supply side of capital comes from the users. Once these vaults are deployed, anyone can deposit funds and benefit from them, in exchange, of course, for sharing some of the profits with the vault curators. These vaults will be completely permissionless, so anyone can deploy one. But I just wanted to give you some perspective on who will manage them and who our first customers will be. There are also asset managers and hedge funds. We're in talks with centralized finance (CeFi) entities that manage billions of dollars and simply want to automate their deployment systems. Quantitative analysts (quants) are extremely expensive and hard to find. They can outsource all of that to our agents and deploy complex trading strategies very quickly, becoming a hedge fund in a week or even a few days. I also want to mention an important point here. As a user, you will be able to stake tokens in a contract very similar to the VEX contract. So you can vote for your favorite vault, vote for your favorite DAO, or vote your assets to increase your rewards when you deposit into a vault. Our product is very complex. The supply side of the vault will be provided by professional users, but the supply of capital will be open to everyone.
• Sam (Cambrian Network): Colin just mentioned allocating capital between yield-generating vaults and lending protocols. Our focus is on measuring where those yields are being generated. To optimize the kind of strategies Colin mentioned, you need to understand the historical yields generated on different chains and across different protocols within those chains. This requires complex data plumbing, tracking on-chain activity across both EVM and non-EVM chains, and tracking protocols within those chains. Builders need both historical information to adjust their strategies and real-time information to execute them. This is one of our areas of expertise—tracking all of this information. If you think about an RPC provider, they provide real-time raw information, which means that the information coming out of the RPC provider isn't always clear and concise. What we do is decode all of that historical data and, based on our understanding of the protocols, we can decode that information and start tracking, for example, the yields being generated. We're currently in closed beta and working with the Coinbase Developer Platform. We're also working with Olas as part of the Olas Hedge Fund Cluster, providing historical and real-time on-chain and off-chain data to agents within Olas. We're also collaborating with several other projects: We work closely with Truflation, providing them with sentiment analysis and wallet activity. Another even more interesting project we're collaborating with is called AskPire. They're tracking tens of thousands of GitHub repositories related to tokenized projects. We track historical contributions and the quality of contributors, and AskPire is building customized trading strategies using our data, allowing them to correlate project activity with future token prices. So, I hope this gives you an idea of the type of information we provide. It's all based on common needs we see among proxy finance projects.
• Renç (Giza): To set the stage a bit, at Giza, we're not really interested in incremental improvements. I think DeFi has been stuck in a cycle of incremental improvements for a long time. What we want to achieve is a complete paradigm shift in user experience (UX) in finance—not just Web 3, but the entire financial industry. We want to shift finance from a product-centric approach to a user-centric approach. We have a very firm view on this point: personalized finance is the way forward. Our vision isn't to create another DeFi protocol, never has been. Rather, it's to create a 24/7 companion that executes and provides insight into your financial situation, helping you achieve your financial goals. This is our North Star. Given the robustness of our infrastructure and this North Star of tailored personalized finance, Giza's agents are now able to serve both retail and institutional users. The institutions we work with today have more rigorous and complex needs, ranging from custody requirements to risk frameworks to liquidity mandates. Giza was built to meet these needs with tailored agent strategies, not off-the-shelf products. This ranges from designing custom agents to providing segregated infrastructure, real-time monitoring, audit trails, and white-label implementations for funds, fintech partners, and neobanks (who have many proactive needs). For individual users, I think there's still room for exploration in this area. This is where we might still offer the same level of sophistication, but with less complexity. For retail users—the "banking the unbanked"—we can enable them to interact with decentralized finance through a radically simplified interface that completely abstracts away the policy layer. We take on the responsibility of financial decision-making for the user. We automate the decision-making process. I think this is one of Giza's most distinct differentiators, and we have the courage, expertise, and talent to take on this challenging task. We are exploring unique requirements for both the retail and institutional market segments. Simply put, retail users want radical abstraction and accessibility, while institutions demand higher standards of security, monitoring, and reporting. We are well-positioned to meet both. Giza has been building a crucial asset base: the stablecoin market. Clearly, it's not going away anytime soon. Its total market capitalization has reached $300 billion, and every stablecoin in circulation represents potential capital that can be autonomously optimized by Giza agents. That's why we built our first proxy for this space and will continue to expand its coverage and capabilities. Of course, this also allows us to serve treasuries, DAOs, institutional funds - anywhere DeFi can be abstracted, anywhere someone asks "How do I invest in stablecoins?" Giza is there.
• Colin (Makina): Renç just said something very interesting about how we can move beyond incremental change in finance. I think all of us are involved in this technology because we recognize that the traditional financial system isn't working for people right now. I think that's one of the guiding lights for everyone involved in Ethereum. What we strive to do at Makina is build safety and security into the foundation of everything we do, while also making it scalable. We firmly believe that by providing this infrastructure, we can deliver the best outcomes for everyone, whether it's a large institution or a small retail user. We view the world very similarly to what Neo is saying. There are entities that have investment needs, and there are entities that want to meet those needs. We're working to ensure that the best managers of financial outcomes have access to tools that allow them to do so safely. We firmly believe this is a growing area. If we look at traditional financial markets, there are currently about $150 trillion in assets under management globally. One very interesting fact is that currently, about 60-70% of that is actively managed, and that percentage has been declining. A lot of this is because people are paying a lot of fees and aren't necessarily outperforming ETFs. We're hearing a lot about ETFs in the crypto world as well. ETFs are also quite revolutionary, changing the perspectives of many in traditional finance due to their low costs. We firmly believe that as we see advancements in security and automation with technologies like Ethereum, the EVM, and AI, these costs can be reduced, allowing people to earn excess returns more cost-effectively through better strategies. This is truly important to us on a global scale. It's not just about doing things better for a certain entity on Wall Street or in the City of London. It's about ensuring that anyone who needs access to financial outcomes can do so. Furthermore, we firmly believe that this should be built directly into DeFi protocols. We should build tools that allow managers to convert these productive assets into collateral or use them in various forms within the DeFi ecosystem. This is how we truly grow the DeFi economy. This could be stablecoins, but it could also go far beyond that, allowing people to match their future liabilities with existing assets and pass them down through generations. We believe this will fundamentally change how people achieve prosperity through their wealth. I repeat, we are huge fans of AI and Ethereum in achieving this goal.

4. In your technology stack, which parts rely more heavily on AI capabilities and which less, and why? Furthermore, since we're discussing building a financial system that leverages AI, risk management and control are crucial. When you consider AI security, how do you factor risk management or control into your workflow?

• Ray (OKX Ventures): When people discuss DeFi agents, it seems that many users, especially retail users, still misunderstand the concept. They may think, "Hey, we can just use AI agents and rely on them 100% to make financial decisions, manage our funds, and search for alpha," but this is actually not the case. In fact, we want to help clients build financial systems that leverage AI capabilities to some extent, ultimately improving work efficiency or decision quality. However, we still need to build a reliable, deterministic, or verifiable workflow because before we invest a lot of money, we need a reliable financial system. This is why I wanted to ask some questions about how you consider potential risk factors in your system.
• Renç (Giza): Yes, absolutely. I think it's crucial. As a company pioneering an agent that makes financial decisions on behalf of users and institutions, this has been one of the biggest challenges we've had to overcome over the past few months—educating the public about the legitimate questions you raise. Will the agent take my money and run? Can we explain what the AI is doing with our money? How certain is that? Are they hallucinating? We had to go through all of this to get people on board. Because it's a completely new tool. In Giza's case, it's not a vault where people are used to storing their money; it's something completely new. Each user has a dedicated agent that serves them. In your question, it's important to distinguish or define what "AI" is in this context. Most of the questions you raise stem from understanding LLM. For us, LLM has been amazing at parsing the common needs of our users and parameterizing them into preferences—essentially translating vague, human-level inputs, from "I want to earn money safely on my stablecoin" to "I want to beat US inflation by 5%" or "I want a moderate exposure to ETH," into structured financial parameters. But the commonly understood aspect of AI—the LLM—stops there for us. Once intent is parameterized, execution shifts to specialized agents built on algorithmic logic and optimization functions that are deterministic, verifiable, auditable, and capable of continuously self-adjusting across markets and protocols. So, by combining these two, we get the extreme customizability of AI with the professional, robust, secure, and policy-bound execution of specialized agents excelling in financial markets.
• Neo (Almanak): I'm not sure how AI can be both deterministic and verifiable, but here's the deal: how we use AI and how we address security. Again, we're very pragmatic about everything. We don't want to reinvent the wheel; we simply adopt what's proven and in demand in the market. We specifically use AI to generate code, which is 100 times faster, and that code is deterministic and verifiable. If you ask any hedge fund manager, any quantitative analyst, any developer if they're familiar with what our agents generate, they will be. If they get a call from their limited partners (LPs) or a bank saying, "Hey, can you show me the code?" they can show it. If money is lost, you'll be able to tell who stole it and how it was stolen, because there's a bug or something else in the code. So that's very important. The security is as secure as any other hedge fund or any other bank. When it comes to execution and ideation, we use agents in very similar ways to everyone else here: basically screening the market, looking for alpha, finding the best solution, finding the best trade, simulating the strategy, backtesting, and simulated trading to avoid slippage. So the AI works with you to ideate, but ultimately you make the decisions. You decide whether to implement the strategies the AI provides you; you decide whether to update the code. The code is fully verifiable and deterministic. Again, we simply took proven approaches and made them 100 times faster in terms of coding and a billion times faster in terms of inference. When it came to the blockchain infrastructure layer, we didn't want to reinvent the wheel; we simply took what worked. We wanted everything to be fully composable, so we use composable vaults. Security is achieved through transparent permissions. Each vault has transparent permissions, so you can see on-chain what a vault has access to. Whenever someone—the vault manager or curator—changes these permissions, it's visible and transparent. This is a perfect replica of what hedge funds do. Furthermore, we created—I think this is very unobtrusive, but one of the most technically demanding and difficult things we've done—a structured workflow for our agents. We currently have 18 agents; seven of them are now available to everyone. These agents are like quantitative analysts, but they operate on infrastructure similar to traditional hedge funds. We borrowed from traditional hedge funds—the infrastructure for creating, backtesting, simulating, and optimizing strategies—but we built it for AI, not humans. Even the creation process itself is as rigorous as any other hedge fund. Essentially, we only use AI for areas that aren't critical to losing money. This gives people confidence to deposit their money, and we've received numerous proactive requests from funds and asset managers to use the tool. I'd say our security is as secure as blockchain.
• Colin (Makina): I've heard a lot of really interesting things from these folks. I probably can't argue the definitive nature of AI, so I'll leave that to you. Again, we're approaching this from a financial perspective. To answer the first question about where we're using AI, I want to emphasize that this is how we're using it today . I'm not an expert in building the inner workings of these AI agents; we strive to provide tools for those experts. We're seeing an evolution. Anyone here who has used AI in any meaningful sense has seen tremendous progress in a very short period of time, and that progress will continue. Where we're really relying on AI right now is in the use of automation. Of course, as we've heard from Renç and Neo, you need to have guardrails in place for this. Makina, one really interesting point is that we're bringing these guardrails to cross-chain transactions. L2 is an essential component of Ethereum, as is L1, the EVM's replacement, and we maintain the same controls when transferring assets across chains. This means we can open up new areas for investment and absorb a lot of information. Sam mentioned some great different sources of information his company provides. Being able to read what's happening on social media—I mean, we're all on X (Twitter) right now—is really key. Most of us have probably spent some time thinking about the Monad this week, consciously or unconsciously. A lot can happen on the Monad, and getting in early can help some people outperform others. But you shouldn't do that without controls. That's what we're really introducing. We believe AI will play a significant role in deciding when and where to deploy capital, but not without controls.
• We firmly believe that these controls still need checks and balances at this point in time. We have a "risk manager" role within the vault. What this really means is that someone can decide, based on a whitelist, what an operator—be it an agent or another person—is allowed to access. This is encrypted and stored in every blockchain activated on our machines or within the vault. So when an operator makes a decision, the timing, direction, magnitude—all of these can be determined in various ways by those operators. But limiting access is something that requires more consideration, and we need to maintain that control. We specifically allow risk managers to use AI tools for rapid iteration and build what we call blueprints or scripts, but ultimately, humans still make the final decisions. On the other hand, from a user perspective, we're running a lot of experiments to better understand what recommendations our depositors want. This is less about executing decisions and more about understanding what users are trying to achieve and helping them align with what's available. As I said, I've worked in traditional finance, and anyone who has spent time there truly understands how difficult it is to access information. We want to help people get the information they want and better understand how they're performing based on their own intuition and their own goals. We think AI is a really good tool for this. It's not perfect. We've been experimenting with our FAQs, which are run through an LLM bot. The team will tell you it still needs more tuning and more data input. But our users have been very appreciative of it, and this has helped us fine-tune our own user interface and front-end experience to better serve these users, allowing them to find out exactly what they want to know in a very efficient way without having to read an 18-page FAQ. Another thing I really want to emphasize is that we don't use AI to write our smart contracts. We have truly top-notch Solidity developers. We're doing a lot of work with auditors to ensure everything is secure under the hood. We believe that, at this point, all of this should be done by very experienced humans, and we're very happy to have experienced humans on our team.
• Sam (Cambrian Network): At a high level, I'd like to share how I view agents. I believe intelligence exists on a spectrum. The agents you see deployed today are what I would classify as algorithmic agents. Their decision-making strategies are deterministic; they're mathematical, use optimization, and behave exactly as their creators intended. On the other end of the spectrum, we have AI agents. The most advanced AI we have today are LLMs. LLMs are creative and adaptable to different conditions. However, the problem we face with AI agents today is that they are non-deterministic. You can give GPT the same prompt and get a different answer every time you run it. In addition to being non-deterministic, they often make mistakes; they can hallucinate. The promise of AI agents lies in their adaptability, which is far superior to algorithmic agents. I believe we will see—and I'm very confident that—the determinism problem with LLMs will be solved. For example, there's a company called Sakana AI, which spun out of Google Brain; they recently published a result that shows significant progress in getting LLMs to generate the same content every time. I believe EigenLayer will also be releasing some similar work. In terms of improving accuracy and illusion, you can assume that the error rate on any nontrivial task will be halved every year. So to summarize, right now, as Renç said, LLMs are very good at capturing intent and translating it into parameters that can be fed into algorithmic agents and then run reliably. At the other end of the AI spectrum, you can assume that their performance will double every year, and they will become active decision-makers in managing our financial decisions. Now, specifically regarding what Cambrian is doing, I'm very concerned about data. Regarding the data issue, we ensure accuracy by using cryptography to check that all our inputs are correct. If you start trying to get blockchain data, you'll find that it's often wrong. Cryptography is the solution to ensure it's correct. This raw data then goes into a database, and when we start tracking things like returns, we have to ensure that our return tracking algorithm is consistent with how the smart contracts are written for all the protocols we track. So we have to do a lot of spot checking and extensive testing with other sources.

OKX Ventures' Thesis on DeFi Autonomous Agents

• Ray (OKX Ventures):

According to our previous research, the DeFi Agent market experienced a critical transition from conceptual enthusiasm to reality in the second half of 2024. The first wave, centered around the "GPT Wrappers/Chatbots" model, promised users could easily navigate complex DeFi operations using only natural language. However, this seemingly promising vision quickly exposed its fundamental flaws in practice.

These early "DeFAI terminals" generally encountered three major difficulties in actual application: first, LLMs struggled to accurately identify highly complex and personalized user intentions in financial scenarios; second, the industry lacked supporting tools to stably convert ambiguous intentions into precise on-chain operations; finally, even if users had powerful tools, they themselves often fell into "decision-making paralysis" where they didn't know what instructions to issue.

However, the deep common root of these problems is that the first generation of agents tried to rely entirely on non-deterministic LLM to dominate the entire process from intention understanding to transaction execution .

This fundamental paradigm flaw led to a rapid market reshuffle. Faced with extremely low actual conversion rates and poor user experience, the vast majority of projects disappeared. The survivors showed a clear differentiation in their paths:

• Some projects attempted to make incremental improvements at the UI level and optimize the prompt word engineering, but this did not address the core issue.
• The other part, which are the projects that are truly leading the market, have chosen a more radical transformation. They no longer force AI to directly understand everything, but instead turn to "autonomous agents" that focus on specific scenarios and provide clear value to users through pre-built workflows.

This emerging class of autonomous agents, through pre-defined, proven processes and focused on building truly deep capabilities on top of the DeFi Adapter Layer and Cognitive Engine, has clearly shifted market focus towards the latter, ushering in the era of autonomous agents. To understand the nature of this shift, we must first clarify the fundamental differences between the two execution paradigms.

We believe that a secure, reliable, and scalable AI-powered financial solution must move away from the straightforward LLM execution model and toward a structured workflow centered on determinism—the requirement that, for any given input, the system always produces the exact same output. This is like a mathematical formula or a piece of traditional computer code, whose behavior is predictable, verifiable, and reproducible . This workflow should adhere to four core principles:

1. Curated Data Sourcing & Environmental Isolation: Agents must access external information (such as market conditions and on-chain data) through strictly vetted and formatted API connectors, rather than simply scraping it from the open internet. This eliminates security risks caused by data contamination at the source.
2. Pre-vetted strategies, not ad-hoc decisions: No trading logic can be improvised by AI. Each strategy must be developed in a sandbox environment and rigorously backtested and simulated before deployment. Its objectives and behavioral boundaries are solidified before entering live trading, ensuring its behavior meets expectations.
3. Permissioned Execution & Risk Boundaries: Policy execution permissions should be strictly limited. Smart contracts should establish clear boundaries of authority and responsibility (e.g., limiting interaction to whitelisted protocols, strict fund transfer restrictions, etc.) to ensure that even in the worst-case scenario, potential losses are kept within manageable limits.
4. Continuous Monitoring & Circuit Breakers: After a strategy goes live, it must be monitored in real time by a 24/7 autonomous risk management system. If the strategy's behavior deviates from expectations or the market experiences extreme volatility, the system should be able to immediately trigger a circuit breaker, taking intervention measures such as reducing positions or suspending the strategy, acting as a final safety valve.

Precisely because of the paradigm flaws of the first generation of products, the market rapidly reshuffled. Faced with poor user experience and extremely low conversion rates, the vast majority of projects vanished. The remaining projects showed a clear divergence in their respective paths: some projects remained focused on incremental UI improvements, while the truly leading the way were those choosing a radical transformation: autonomous agents . Don't misunderstand this concept: these emerging agentic products no longer require AI to understand and perform everything. Instead, they provide clear value to users in specific scenarios through pre-built, validated workflows. Their R&D focus is on building a truly competitive DeFi adapter layer and cognitive engine. Consequently, market focus has clearly shifted to the latter, ushering in the era of autonomous agents.

Conclusion: Despite widespread skepticism surrounding the Crypto-X AI sector, we firmly believe that by adhering to the aforementioned principles and properly leveraging LLM capabilities, this field can offer a compelling value proposition, particularly to institutional clients. This includes enhanced multi-dimensional information analysis capabilities (capturing complex factor correlations that traditional algorithms struggle to capture), orders of magnitude improvements in code development and deployment efficiency , and even more robust automated execution capabilities . Therefore, we are committed to continuously monitoring the development of this field over the long term and seeking out early-stage teams that align with our core principles.