An Agent of Record: Chris Zhu on Building Donut's Personal Trading Agent

0:00Josh Han Welcome back to the content suite in Singapore. We are continuing our discussions on AI agents and the frontier of trading. Today, I am joined by Chris Zhu, the founder of Donut AI. Before we get into Donut and what you're working on, I would love to learn a little bit more about your journey within crypto.

0:17Chris Zhu Yeah, for sure. So joined the industry 2021 with a couple of friends. We're mining out of our dorms very early on in college, and then they're on join a company called ByteDance, or TikTok's mother company. Was there doing products and really fell in love with crypto late 2021, early 2022. Met the early FTX squad, the Galaxy squad, and was lucky enough also, you know, kind of volatile enough to get their funding very early on, and started building my first company, Sonic SVM, which was basically building our own chain infrastructure for vertical trading apps as well as gaming. And that really kind of took off in 2025. So I've been running that with my co-founder and he's taking on a lot of the daily operations for me to think about what that next step looks like for our products.

1:02Chris Zhu And earlier last year, I fell in love with the idea of just using AI to browse information and how that human traction and agent interaction will look like and start building Donut, which started out as an AI browser company, and later on evolved into a trading personal agent. We'll talk about that.

1:19Josh Han Yeah, I mean, so Chris and I have previously worked together and while building Sonic, I know that you had first person experience basically working on a lot of different apps. And I think a lot of those insights, whether different models through TikTok led you to building on Donut. The first iteration of Donut was supposed to be a browser. Yes. Right? But since then, you know, I think at that point, we're probably still on ChatGPT-5. I'm assuming new models had not come out and we have the kind of introduction of new harnesses,

1:50Chris Zhu right? Right, right. So tell me a little bit about how you guys went from browser to agent now. Right, so we started building Donut probably second half of last year. And back then, it was the only agent product out there which really took the form of a cloud computer and giving agent access to your cloud computer was really Manus. Manus was kind of like the first to put that flag over there. And as soon as Manus kind of came out, we saw the possibility of having kind of separate expertise for these agents. So we really believe that for trading agents, we needed to have a different set of automation skills. So for example, OpenClaw, when you set a cron job, defaults to a 60 second cron job activation which then also uses quite a bit of capacity and resend box. So we think that for trading, for example, you need subsecond and really, really fast automations.

2:40Chris Zhu So you need WebSocket connections rather than having like the default OpenClaw ones. So a lot of these specifications kind of came about when we were really playing around with OpenClaw earlier this year, so around February or March. And then we decided to kind of fork that and build our own harnesses similar to Muse did, really for general agents. But for trading, we needed to do that for automations, for strategies, for the paper trading, such and so forth. And we gave, we really think about it as giving the agent the ability or more capacity to do anything on chain because money is going to move on chain and all of the money is going to be delegated by agents as well. So that's kind of how we started playing around with the idea of building our own harness. And that was the early days.

3:23Josh Han That makes a lot of sense. I think for people who are getting into agents, a general agent like Muse has been really helpful. There could be obvious potential downsides like giving Mark Zuckerberg your data, but other than that, what does an agent need to be your financial agent or your on-chain trading agent that maybe Muse or Instinct can't

3:41Chris Zhu provide immediately? Right, right, for sure. So if we look at Muse or Instinct, their business model is first and foremost, it's supposed to be a general agent that's open to everybody. But that comes with downsides of the model capacity as well as whether you can use SOTA or state-of-the-art models to solve for complex tasks like research for a token drop or calculating valuations for a project you want to invest in. Those are capacities that require the best models. And on top of that, we really want the models to be precise at their execution. So they cannot just hallucinate off of a mathematical test. I can't even afford a 1% failure rate there. So we needed to build all of that execution stack and then connect it to a model.

4:27Chris Zhu So the model is just interpreting the intent and then filing kind of a format, which is verifiable, and then giving the user the ability to prove that transaction every single time. So we need that to be verified on-chain and off-chain, and also have that entire log of data shown to the users. So really, we think of a personal trading agent as kind of an agent of record. And if you look at Muse, it really doesn't have to show you every single step of all of those logs. But for trading agent, we really need to show that to the user and the portfolio, the P&L. Every single step, that change needs to be visible to the user as well. So I think kind of drawbacks there is that we have to spend a lot of time thinking about the UX and UI and really showing how much money has changed over the course of the usage of the agent.

5:21Chris Zhu Very cool.

5:22Josh Han So with the UX becoming more

5:24Chris Zhu mathematically correct or just financially

5:26Josh Han accurate and less hallucinations, what do you imagine, let's say, a couple of years down the line, if you're talking to your donut personal agent, what do you imagine the best experience to be? Let's say, is it executing on behalf of you like a brokerage? Is it giving you financial advice about your overall portfolio? What's your dream scenario for an agent?

5:48Chris Zhu Yeah. So when we look at any agent product, we think the first step is giving the agent the full context of what you already have. So whether it be Grokbot, when you connect it to your Slack MCP, your Linear, for productivity suite, you really need all of those external connectors. Similarly, in trading or even for personal finance, we think that you should start with those connectors and giving the agent context, which means we have to be Switzerland for all of the exchanges instead of building our own perp DEX or putting our own trades off-chain. So what that means is if you are already trading on Hyperliquid, you should be able to connect to Hyperliquid your record of trades very, very easily. And simultaneously, if you're landing on Jupiter, then you should be able to connect to that as well.

6:33Chris Zhu So every single protocol then becomes a connector on top of Donut, and that's kind of the marketplace for plugins, specifically for on-chain finance and personal trading. On top of that, these kind of connectors also provide liquidities. So Jupiter, for example, has their own liquidity venues, and we can talk to ChainUp or all of these other infrastructures. So they became kind of a venue for us. And if the user is already trading on those venues, then it's basically their context. If they're not, then the agent has this new capacity that you can connect into or tap into. So several years down the line, the ideal experience should be that Donut has access or has knowledge of all of your portfolio information, gives you trade advices so that you can rebalance between Jupiter landing and Aave regardless of the chain.

7:23Chris Zhu And then if you have an opportunity that is based on a thesis, for example, I personally believe that tennis is going to be huge. Yeah, right. And off of that thesis, we've seen products like Supertake, for example, incubated by USV that centers around that thesis. But how do we find the right opportunity of on-chain assets to then invest into that entire process, should be abstracted away by the agent, off of the money that you've already deposited into Donut. Yeah, so that's kind of the ideal experience.

7:56Josh Han That makes a lot of sense. I mean, that would be a role that you would have to basically

7:59Chris Zhu hire an analyst for, right?

8:02Josh Han If you're interested in a general category, but you want to do more research on how

8:05Chris Zhu do I get higher leverage or higher beta exposure to this,

8:09Josh Han or what are the newer companies that are coming public within a specific category or topic? Very cool. One thing that I am personally concerned about and would love to learn more is with personal agents, if ChatGPT Astra is already listening sometimes with my answers, I know that you are focused more on giving accurate portfolio information or trade construction, but how can I trust it with my private keys? What are some of the standards maybe you guys are creating or standards in general, that you see happening where agents in general have more financial access and they might not get prompt injected or hallucinate and send money where you

8:48Chris Zhu don't want it to go? Yeah, like we said before, the execution stack and the model intent passage are separate. So similar to kind of Muse, where if you look at their stack, there's something called a Sentinel, which is kind of this middle office or this middle stack that takes care of all of the authorization. The model itself does not have access to the authorization. The users have to very explicitly state the intent of authorizing the specific action. Similar to Codex, that's how we've been able to kind of build our MCPs across the stack. But at the same time, we think that as models get smarter and smarter, those hallucinations will decrease and we will also basically build an eval suite of all of the mistakes the agents have made and make sure they don't make that again.

9:39Chris Zhu So that stack of data of mistakes will grow as we have more and more users as well. So right now, really, it's to separate the authorization from the model to the user's explicit intent and then really build up that entire kind of eval loop.

9:57Josh Han I'm very curious. I'm sure you and your team have tested different models with financial data more than the average person. Between Claude, ChatGPT, and the OpenAI models, maybe there's some open source models that might be more helpful. How do you view which models are best for, let's say, a financial agent? And say if someone wants to start off, maybe not with an agent, but just giving their financial data to a model, how would you recommend that setup?

10:24Chris Zhu Yeah, for sure. I think each company and each product, depending on the target users, have a very different answer for this question. The main reason is it's really a trade off between margins and effects or quality and speed to a certain extent. The latency is more so like a scaled after effect. What I mean is that, for example, for our end user, we want to serve the everyday consumer, similar to the target user of FOMO's. They are onboarding to crypto and the total pool is very, very large. If we look at the user pool of Muse, it's actually only a couple million right now. But the total chat pool of Facebook's products is around over three billion total users. So we want the pool to be extremely big, which means that we have to offer an affordable model for most of these users.

11:15Chris Zhu Grokba, on the other hand, are for users who's paid for either X or Cursor's premium subscription. So they can afford to cover the cost of the SOTA models like Musk just tweeted today. So when we think about it, first and foremost, we want it to be very, very high quality, but also affordable for the everyday user. Which means that if we look at the total token consumption across all of the models, DeepSeek actually has been consumed more than all codecs, Claude and the other models combined. The reason is the cost to affect ratio is so high that you can just use 4.1 and have a generally acceptable model. On mathematical hallucinations, on the other hand, GPT has performed by far the best across all of the models that we've developed.

12:08Chris Zhu So Luna is a really, really good trade off. We're just looking at mathematical hallucinations and the accuracy rate. But you cannot prevent Luna from hallucinating if we're looking at the percentage of chances. It's probably anywhere between 95 to 100% accurate, but that's not enough for trading. So the next step is going back to separating the model from the users checking the hallucinations.

12:35Josh Han That's very cool. I mean, I think there's a lot of alpha actually being dropped because we've heard different models. Surf was on here and they're using a lot of Grok. And I think that's mostly for cost saving and Twitter. GTE was using Opus 5.5. I'm very curious if we could almost have like a trading agent competition where we set it up, give each all the bots like $10,000 and see what happens. But that's very cool. Tell us a little bit more about what you're excited in terms of Donut, where you see the next couple of months going and if there's anything

13:06Chris Zhu else that you want to share. Yeah, one thing that we're really excited about is really just going to the extreme of building a really, really good mobile agent. So we've seen Muse as well as Instinct kind of going off of messaging platforms for really evolving to a couple of different kind of verticals. Instinct profits much off of the entire kind of travel space or hospitality space, having around $1 billion in transaction volume across the board, but having a free agent that lives off of iMessage. So a lot of their new raise is going really into the inference cost. So really planning for compute. And I think similarly, we have a very similar business model when it comes to trading agents because we want to give this model or give this agent for free to the end user. But they will be trading and really racking up volume as the agent becomes better and better.

13:55Chris Zhu So what that means is in the next couple of months, we need to make that mobile experience really, really good. So when we look at some of our inspirations, FOMO is one where it creates this incentive for users to look at the social groups that they are invested into, whether it be a clan or following your influencers and really finding targets there. But for agent users, our product is really one, the marketplace of connectors, having gold assets, having XAU assets such and so forth with different kind of venues. And then second is to directly follow other people's agent trades where they have automated trades and strategies already delegated to the platform.

14:41Chris Zhu So those are going to be two of the biggest features coming up in the next month or so.

14:46Josh Han Nice. So if I have an agent with a specific strategy, maybe I can put it up on the marketplace and then someone can use my strategy and they can manage a portfolio. Very cool. Going back into Instinct, since you guys are both creating general models, tell me a little bit more about what is the business model if the inference is free?

15:08Chris Zhu Excuse me. So if the inference or at least the base model itself is provided for free to the user, the business then becomes making the trust layer of that agent heavy enough so that the user is willing to put money into the agent or delegated to the agent. So there's two different business models when that happens. One is more like a brokerage where you are really transmitting traffic to Binance or some of the other exchanges like Hyperliquid. On the other hand, there is also the custody model where we are building off of Privy or Turnkey in this case and really having all of the deposit coming to Donut. And then you trade directly at Polymarket or Hyperliquid and then we have that builder fee. In both of these cases, we're taking off of a take rate off of these already existing liquidity venues.

15:57Chris Zhu And we can imagine a day where we build our own perp DEX or our own prediction markets, whereas our end consumer is racking up the volume. But that's going to be much later down the line, similar to what Pump did, for example, with PumpSwap. But in our business model, really, it's about whether we have a empty wallet for the users to start with their kind of agentic trades or they transmit their existing context into a existing or take their existing wallet and give agent access to that. So that's how we think about the two different business models. Very cool.

16:33Josh Han That's actually I've never thought about it like that. Do you think that there's correlations with, let's say, social media being free? I would say my concern is if Instinct or a model, a business model like this is like maybe my data could be used because maybe my trades are all unprofitable or like there's just not that much. But there is something with data privacy. If you could speak a little bit about how you view your privacy, because I'm sure you're using a lot of different products. If you have any concerns, there are maybe ways that you're protecting yourself from

17:05Chris Zhu worst case scenarios. Right. And we've seen all kinds of users from institutions to kind of individual traders. Everyone has really, really different preferences on this. But I think a general kind of take is that when the agent gets its own computer, not a lot of users are willing to delegate all of their authorization to that new computer. But if your workflow lives locally like Codex, for example, then the agent gets access to a sandbox browser on top of your Mac or your local environment, then you're much more willing to give all of the access to the agent. So we think from a personal agent perspective, we need to give the agent both of those capabilities, whether it's a cloud computer or a local environment.

17:50Chris Zhu And it's really the user's choice of do you want to set up all of your accounts on the cloud computer or just delegate that ephemeral state to the local environment.

18:01Josh Han Very cool. It makes a lot of sense. I mean, I'm sure you've been enjoying Singapore. We have MBS in the background. I'm curious if there's any other trends. You're always on the frontier of things ever since I've met you that excite you right now in crypto other than the AI agent space. Yeah. So we've been spending quite a bit

18:22Chris Zhu of time in the States and in New York in the past year or so. We've realized there is a huge opportunity still, I think, in just the perp space. And a lot of the perps products have gone more so into regulated territory. And some of our most bullish ideas have come from some of the new regulations being discussed within the US. And I think similarly for agents or trading, that's going to be a big topic of how do you delegate access to money directly to these agents? And what is the on-chain version of that? What is the regulated version of that? And we've been kind of talking to a lot of authorities in the States about that as well. So that's one thing I think that's going to separate different products being built.

19:07Chris Zhu And then secondly, then the big user of FOMO myself. So just looking a lot at the trends between having kind of a underlying asset tied in with a meme coin. So having that 1000X stream while having something like AI pairing token or Nvidia pairing token like AI. And on that route, I think we can take it further to kind of anagolize the IPO market and taking some of those equity entities and launching a token off of that. That's going to, I think, evolve very, very quickly as well. Yeah.

19:46Josh Han That's very interesting. I mean, on that note between FOMO, even between you and I, we've discussed a lot of the social trading platforms. FOMO is not the first one to introduce it. They might even have a more limited feature set without like limit orders and such. I'm curious actually to get your opinion. If you're a founder now in crypto, what do you think is like a secret sauce or maybe something that FOMO did correct that you would want people who are listening or other founders that you talk to to try to replicate? Or from your multiple times as like a crypto founder, I'm curious, what do you have to do in this space to win compared

20:22Chris Zhu to other industries? Yeah. So full disclosure was an early investor into FOMO, but I have no insider information of how they build the products. But I was really bullish Vector before FOMO. So that was kind of the Tensor team building the first mobile trading product on Solana. Yes, exactly. And I think FOMO took it one step further. Tensor was entertaining the idea of trades being one, social and two, entertainment. So they had a vertical scrolling feed that looks like TikTok for each one of the charts. FOMO, on the other hand, focuses more on the social aspect and then having a information feed that looks more so like a vertical feed just for trades. But it's just text. But if you look at the tax for FOMO, it's very much there's only thesis and trades.

21:08Chris Zhu And for trades, you only have long or short. So it's extremely simple for the end consumer. And even so, we probably have like two million total registered users on top of FOMO. So that end pull compared to three hundred million off of Binance is a tiny, tiny percentage. So I think I always say like every generation, there's I think one generation crypto is like four years. Every generation has their own exchange. And FOMO is the next exchange, almost like that format. But I think as a founder myself, to replicate that success, one thing that I think FOMO does very much right, aside from just kind of the product thesis, is how much time they spend incubating as well as finding traders very, very early on. So they're in the show and around having I think it's like 100, over 100, 140 traders and founders.

21:58Chris Zhu And then secondly, really having diamond hands on top of the leaderboard. It's a social product at the end of the day. So I think a lot of those traders as well as the meme devs are very close and very early in the product journey. And they really understood their users there. But that's heavy work. Yeah. So a lot of respect to the team there.

22:23Josh Han For sure. I mean, I totally agree. How do you get as much buy in at

22:26Chris Zhu the beginning as possible?

22:28Josh Han I think we've had this similar discussion when we talked with Ethena earlier today of like they really won the BD game and talking to the exchanges, getting them to build the stablecoin product out is really was really BD was the secret sauce. I'm going to leave it up to you. If there's anything else that you want to share regarding Donut or

22:48Chris Zhu anything else, please feel free. Yeah. Well, I think I just want the listeners to know that Donut is the personal agent for on-chain trading. We'll start with on-chain trading and go into all of finance and connecting to every market, whether it be Hyperliquid finance, Polymarket. And we really do not discriminate here. We think that the connector as well as the leaderboard are the agent trading products that we're building. And the agent really facilitates both of those products. Yeah. So thank you for having me. Absolutely. Thank you, Chris, for coming on.

23:20Josh Han Appreciate you. Nice.