Podcast: The Round Trip

Compiled & edited by: Yuliya, PANews

This was an in-depth conversation about encrypted data, AI agents, and the future of work paradigms. In Founder's Talk, a new series from PANews and Web3.com Ventures' "The Round Trip," host John Scianna invited Alex Svanevik, CEO of Nansen AI, to discuss how Nansen evolved from a purely on-chain data analytics platform into a full-stack intelligent product driven by AI agents. Alex not only shared the polar explorer spirit behind Nansen's name but also delved into the evolution of the "smart money" label, the "trust ladder" of agent transactions, and how AI is reshaping engineering culture and organizational structures within enterprises at an unprecedented pace.

Nansen AI's original intention and transformation

John: Now is a particularly good time to interview you because AI is really taking off at an accelerated pace. I'd like to ask you to tell us what Nansen AI actually is. Is it an intelligence layer? Or a data company? Or will it eventually become an agent?

Alex: Let's start by talking about the original intention behind Nansen. Our mission is to uncover effective signals and create success stories. In our first six years, we spent a lot of time digging out and refining effective information, and now we're shifting more towards "creating success stories." From a product perspective, our North Star goal is to help investors grow their portfolios more effectively. The measure of success is whether the product can truly help you make more money.

Under this general direction, we have essentially completed a transformation. Originally, Nansen was more like a pure on-chain analytics product, but now it has become a more full-stack on-chain product. Previously, people used Nansen because we had strong on-chain intelligence and analysis capabilities, but now you can stake on Nansen or trade directly on-chain through an agent . I don't like to force products into a single category; the best products often cannot be defined by a single category, but essentially they all help on-chain investors grow their portfolios in unique ways.

John: You've been using the Nansen AI domain, but AI today is completely different from what it was six years ago. You've established a foothold in the data field and continue to expand. What changes do you see in AI over the past six years?

Alex: Actually, I've been working with AI for almost 20 years. I studied AI for my master's degree from 2009 to 2010, and my first job in 2010 was at an AI consulting firm. Back then, not much was happening in the AI ​​field; everyone thought it was science fiction. But things have changed drastically since then, whether it's the past 16 years, the last few years, or even just a few months.

For us, AI has always been present in the backend . As you know, Nansen is probably one of the best teams in the world at address labeling, and this is precisely what we've accomplished with AI. We've labeled over 500 million addresses, and almost all of the labeling work is done algorithmically , with a large portion of that using heuristics.

Now, AI's role at Nansen has changed; it's moved to the forefront . The way you interact with the product has become agent-driven. If you're using our mobile app now, it's an AI that you can ask directly, like chatting with Grok or ChatGPT: "What's smart money doing today?" "Which token should I focus on?" "How's my trading style? Give me feedback, like what I did right, where I lost, and how I can improve." This AI agent is extremely flexible and can even place trades for you.

The Explorer's Spirit and "Smart Money" 2.0: Using AI to Predict Future Winners

John: Would you like to talk about where the name Nansen comes from, and how you understand "him"? After all, what you're doing is essentially exploring on-chain data.

Alex: No problem! Our company is named after a polar explorer named Fridtjof Nansen. He was not only a scientist and explorer, but also a diplomat, politician, and prolific author.

I've always viewed people in the crypto world as pioneers, venturing into uncharted territories. My own on-chain "exploration" has been particularly fascinating, sometimes encountering things no one has ever seen before, such as a specific address, pattern, or investment opportunity . I see our users as explorers (like Fritjof Nansen), and we as their compass, or rather, a toolkit, a vessel, ensuring you safely reach the North Pole.

John: I feel the market is in a "colder" phase right now, and that's a very apt analogy. On-chain data and analytics have changed a lot, and people are starting to use new wallets. How do you use AI to better identify these patterns?

Alex: The crypto industry is constantly evolving, and you have to adapt continuously. The beauty of agent-based user experience lies in its fluidity and adaptability to change. If you create a traditional UI and hardcode narratives like NFTs, memecoins, or Desci and social coins, the product quickly becomes a cobbled-together "Frankenstein's monster." (*Note: Frankenstein is a monster in Mary Shelley's science fiction novels, created from different bodies. In the context of technology and product, it often refers to "Frankenstein's monster" products that lack a unified underlying architecture, blindly chase trends, or forcibly pile on features, ultimately resulting in a highly fragmented user experience.) But agent-based UX can theoretically expand infinitely; it only highlights what is truly important.

The same applies to data; products must be flexible enough. Take "smart money" as an example. Typically, several thousand addresses receive this honor badge, and I've even met people with this label offline in Singapore. However, we've always maintained a vague definition because a definition from five or six years ago may not hold true today. For instance, Three Arrows and Alameda both had the "smart money" label back then, but they both failed later, so it's inappropriate to label them now.

It's difficult to consistently apply a profitable strategy across different timeframes . However, we believe the data contains enough signals to make predictions with a much higher accuracy rate than flipping a coin. We use machine learning and AI to predict which addresses are more likely to make money next week. Building on this, we'll be launching a new product later this year (hopefully in a few weeks), a kind of "smart money 2.0" that essentially predicts which wallets will make money in the future, rather than just telling you who made money in the past.

This analysis also includes determining whether they are agent traders, pure traders, or possess some kind of advantage . We look at dozens of characteristics, such as holding time, number of trades per unit time, win rate, ROI, performance relative to BTC/ETH or an index, and performance relative to a specific index product .

We prefer to feed the model more features so that the AI ​​can figure it out on its own, rather than intervening too much in human intervention, and then evaluate its prediction accuracy.

The future inflection point of agent transactions and the "trust ladder"

John: Yes, I saw you released Nansen's CLI a few weeks ago, about a few weeks ago. And I know one of your predictions is that agent traders will eventually surpass human traders. How soon do you think that inflection point will arrive?

Alex: Currently, most people still manually select tokens. Of course, they might use tools like Nansen, along with some intuition, but I do think that looking at the bigger picture, by 2030, transactions completed by agents will definitely surpass those completed manually by humans. I think this might even happen sooner, perhaps as early as 2028, and 2027 might be the year we see a huge shift.

This is similar to programming evolving from handwritten code to vibe coding, and then to agent-based engineering. We're currently in the manual trading phase, and I predict vibe trading will become more popular this year, while true agent-based trading (allowing agents to trade entirely autonomously on your behalf) will see a significant surge next year. By 2028, even more trading volume will come from agents.

John: Right now, it seems there aren't that many safety barriers; only a very small number of people would use agents to make five-figure transactions.

Alex: Yes, that makes sense. Because there's a "trust ladder" that you have to climb step by step, especially for users. You have to build trust little by little, you have to see results first, and then slowly detach yourself from the (transaction) process.

I believe that as a product provider, we must also provide a similar ladder, instead of letting users put their money in and leave it to chance right from the start. You have to tell users: "You can try this method. You can use our Agent as your 'assistant' to participate in transactions, but ultimately, you are the one who executes each transaction ."

Then you'll gradually realize: "Actually, I don't think I need to click that button myself anymore, so why not just let it do it automatically?"

Even so, you might still want to see the execution trajectory or receive notifications. At a later stage, you might switch to the next mode: giving it a relatively broad strategy direction, and then hoping to have a way to backtest this strategy. This is actually one of the features we plan to launch later this year (a native AI backtesting framework). Only after that will you truly be willing to put it into automated trading mode.

However, this process must be built upon a foundation of trust built up layer by layer. I don't think it's reasonable to jump straight into "automatic" trading. Many companies will give you tools to do this, but many people end up losing a lot of money because these tools are far from mature. Frankly, I think doing so is somewhat irresponsible. This is not what we want to do; we want to do something long-term, so our approach will lean more towards the "trust ladder" path.

John: A lot of people have been trying OpenClaw lately, and then they say, "I'm using it quite smoothly now, so can I give it email permissions as well, or give it other permissions?" Yes, that's actually the trust ladder.

Alex: And I think it's very similar to self-driving cars; the principle is the same. You don't just sit in the back and let it drive on its own right away. You'll let it change lanes first, or only drive in a single lane, or you'll still be in the driver's seat. You're just giving it a little bit of control, even if the speed is a little slower.

Nansen has fully embraced AI internally.

John: So, OpenClaw accelerated the development of your agent trading strategy to some extent?

Alex : First of all, OpenClaw is a milestone in software history, and we use it extensively within our company. In fact, we now have more OpenClaw users than actual employees , so we are heavy users of OpenClaw.

As for the transaction agent, we're using Pi Agent, a product we haven't officially released yet. However, in our early internal experimental versions, we don't need the full OpenClaw stack, so we streamlined it into the lighter Pi Agent because it's more portable. Of course, we're also exploring other solutions.

Overall, however, creating a transaction agent is much easier for most people than before. For example, you can use OpenClaw directly. Of course, if you're only using it for trading, it might be overkill. But generally speaking, I have several OpenClaw instances that are integrated with the Nansen CLI. Because of this, they can create wallets and execute transactions, and can directly take action after identifying trading opportunities they deem worthwhile.

John: How do you operate as an AI-driven company? You were one of the first companies to truly implement OpenClaw on a large scale, climbing the trust ladder very quickly.

Alex: We had already developed a very clear AI strategy by the beginning of 2023. At that time, we realized that AI would evolve at a very, very fast pace. So it's not that we jumped directly from traditional working methods to OpenClaw overnight; it's actually a gradual process along the maturity ladder, which has been going on for more than three years.

I've talked about this on other occasions, about how we made the whole team more AI-native. When we saw what OpenClaw could do, we realized we had to configure it internally. It was a bit of a "prisoner's dilemma" at the time, because many team members would use it regardless of our agreement. The best approach was for us to proactively provide an internally hosted, more secure version of OpenClaw by default. And this version had to be good enough so they wouldn't feel the need to build a crude, cobbled-together, and insecure version of OpenClaw.

We had our security lead and senior engineers work together to build an on-premises deployment platform. We deployed it using isolated virtual machines, used Kubernetes and Helm for configuration management, and prevented instances from updating automatically, in order to achieve the highest level of security possible.

John: Have you noticed that the smartest engineers are the first to use it?

Alex: The reality is more complex. Some engineers who were originally at the mid-level climbed all the way to the top levels because they embraced agent-based engineering methods more quickly. The working methods of Claude Code and Codex give people a feeling of being constantly "stimulated by dopamine," and every feedback is a stimulus.

Here's an interesting paradox: you're actually less likely to use AI in your most familiar areas. Engineers might be more inclined to use AI in design or product management, while designers are more likely to use AI for coding. So sometimes non-engineering roles actually progress faster and are more productive than many engineers who haven't yet embraced AI.

John: How much faster are your delivery speeds now compared to last year?

Alex: Our deployment rate per capita (based on valid PRs merged into production) has increased fourfold in two years. Meanwhile, the change failure rate hasn't worsened; in fact, it has slightly improved. The median change lead time (from ticket creation to deployment) is now below one hour.

John: Is your current process that the agent sees the ticket first, then writes the code, and then a person reviews it?

Alex: Yes, it's an automated triage process. After a bug is found, one agent automatically creates a support ticket, another agent listens and automatically generates a pull request, and then more agents automatically review them. In the Nansen team, pull requests written and reviewed purely by humans are very rare now.

Once you have both the ability to build a closed loop and meet quality standards, human judgment becomes the scarcest resource. Delegate the mundane, basic checks to the agent, while humans (engineers, product managers, or designers) retain the ability to make overall judgments, such as whether the feature itself is meaningful or will have negative impacts.

John : So you guys also have to reorganize the team?

Alex : Yes, small teams are the future . I've worked on several special projects where the team size was only two or three people (such as two engineers, or a researcher/designer plus an engineer). We are moving from "fewer but larger teams" to "more but smaller teams".

Each team has a very clear division of responsibilities, much like a symphony orchestra. There are violin sections, percussion sections, and so on. Everyone focuses on making a specific component of the technology stack (such as address tagging, transaction execution path, authentication system, embedded wallet, agent underlying framework, etc.) the best it can be in the universe. You have to manage your own line of work; the violinist shouldn't worry about what the percussion is doing.

Breakthrough in physical and network bottlenecks

John: These changes have also brought a lot of qualitative data, such as from the use of tools like Auto Research.

Alex: Yes, we're starting to use AutoResearch more to optimize front-end latency. In the agent era, one of the things I valued most was minimizing latency . Agents naturally prefer low-latency products, especially in transactional scenarios. If one product can do 1000 things per unit of time, while another can only do 10, the agent will definitely favor the former.

We've spent a lot of time on this, for example, rewriting certain components using Zig or Rust, which significantly reduces latency. In the AI ​​era, product development requires constant awareness of bottlenecks. AI is accelerating everything that previously lacked obvious bottlenecks, making bottlenecks in the physical world, network latency, regulation, and internal coordination more apparent.

John: Agents communicate with each other too much, and humans process more information every day, so attention has become a bottleneck.

Alex: Absolutely. We currently have about 80 OpenClaws running on Slack, some of them incredibly talkative. When everyone wakes up to a long message from one of these "Lobsters," they don't want to read it at all. So we have to do governance, like streamlining their communication by 50%, and institutionalizing these practices, writing them into the shared company context and culture for all "Lobsters." To be a "good Lobster employee," you also have to adhere to the company's values.

John: How do you think the external operating system of AI will evolve in the future?

Alex: I personally have a very positive view of open source . OpenClaw represents an open-source agent runtime framework, while DeepSeek is a landmark case of open weighted models; their influence is much greater than I imagined.

The open-weight model is already very powerful. Last Friday, we deployed an open-source model (un-tuned version) in Nansen AI's Fast Mode. Its median time for the first response token dropped from nearly 8 seconds to about 4.8 seconds, almost halving the latency and making it 50% faster. We found that the marginal improvements gained through fine-tuning are no longer worthwhile, and the un-tuned version is less burdensome and easier to switch between different service providers. Therefore, we prefer to use the un-tuned version.

John: Thank you so much for coming to our show. AI is not only the future of trading, but also the future of work and Web3.

Alex: You're welcome. I recommend downloading Nansen AI from the App Store or using the web version. If you want to install the CLI, you can simply run `npm install nansen-cli`.