Virtuals Protocol and the intersection of Agentic AI & Web3 with Jansen Teng
Fresh out of the studio, Jansen Teng, co-founder and CEO of Virtuals Protocol joins us to share the mission and vision of Virtuals Protocol and how the protocol intersects generative AI and web3, unlocking the world of autonomous AI agents powered by tokenization to unlock new economic possibilities. Beginning from his origin story, he explained how the Virtuals team explored and eventually come to the realization how tokenized AI agents can power entertainment, gaming and trading with Luna, the AI avatar that grew through the help of AI agents and even inspiring works of real art. Last but not least, Jansen offered his perspectives on the trends of agentic AI and web3 and elucidate what great looks like for the Virtuals Protocol.
"What does great look like? For us, it's being able to achieve this vision of a society of agents. So, if you can show clearly the economic value of a society when agents can influence other agents, agents can influence humans, and humans can influence agents, that would be something that I'll be very proud to say that we've accomplished. Because that's the beauty of really combining breakthroughs on the AI, on the autonomous agent front, and the value add of crypto. This is one of those very rare moments that crypto has that true potential value add, right? As a productivity lever and an economic lever in society." - Jansen Teng
Profile: Jansen Teng (LinkedIn, Twitter), co-founder and core contributor, Virtuals Protocol (Twitter, Substack)
Here is the edited transcript of our conversation:
Bernard Leong: Welcome to Analyse Asia, the premier podcast dedicated to dissecting the pulse of business technology and media in Asia. I am Bernard Leong, and the intersection between generative AI and Web3 creates new opportunities and challenges across different industries. What does it mean for the future of AI agents, digital economies, and human-machine collaboration moving forward? With me today is Jansen Teng, co-founder and CEO of Virtuals Protocol, a pioneering platform at the forefront of the AI Agent Revolution.
I have a disclaimer to make. I have been working with the virtual teams as part of their advisory council since a year ago before they broke out. So I do have some virtual tokens. So here is my disclaimer: no financial advice and not even life advice. That being aside, the Virtuals protocol team has been one of the best teams I've worked with on AI in the region. With all the disclaimers out of the way, Jensen, welcome to the show.
Jansen Teng: Pleasure to be here, sir.
Bernard Leong: Yes, I've listened to both your recent interviews on the famous Bankless podcast and also the one with The Chopping Block team under Laura Shin's Unchained network. Both of us interacted and hung out together during F1 last September. I want my audience to know your origin story. How did you start your career, and what sparked your interest in crypto, gaming, and AI?
Jansen Teng: Yeah, so actually, two parts, right? On the crypto front, it started when I was in my second year at university. I think it all started with 90% of people in crypto — it's like, "How do you make money online?" And then they'll be like, "Oh yeah, mine Bitcoin," right? So it started from that and then snowballed into discovering Ethereum. I started mining Ethereum in my dorms running on free electricity. That was my original exposure to the whole decentralized application space, but honestly, I didn’t do much there.
Meanwhile, on the other side, there’s always been this entrepreneurial itch. I started a deep tech venture back at Imperial as well, but after that, I left and went into management consulting, mainly to pay off my scholarship bond. On the weekends, I was working with my co-founder on other types of startups in the webtoon space. It ranged from digital marketing, and standard stuff, to even starting an AI recommendation platform for property. We tried to take on the likes of PropertyGuru and others, but we quickly realized that sometimes it’s not about the product; it’s about how hard it is to dislodge front-of-mind players, especially in a marketplace or network-effect environment. So we had to move on from that. That was our initial dabble into big data and the ML space from a product standpoint.
Fast forward to 2021 — this was the bull cycle of crypto.
Bernard Leong: They used to call it the Super Cycle.
Jansen Teng: Yeah, it hasn’t reached Super Cycle yet. But that was when my co-founder and I did quite well from an asset management standpoint in the space. We had the luxury of leaving our jobs, so we started saying, "Hey, let’s build full-time in the crypto space." Back then, we had a ton of gaming assets — Web3 gaming assets — so we started a gaming DAO. Think of it as a gaming investment vehicle on-chain. That was our initial venture as capital allocators in the space.
But then 2022 came, and we knew everything blew up, right? To manage risk better, we realized that allocating capital at arm’s length was a very low risk-adjusted return play. So we pivoted to a venture studio model and decided that if we wanted to build in this space, we’d need to recruit our co-founders to build projects at the intersection of what we believed in. Back then, we saw opportunities in consumer-facing applications intersecting with crypto and AI agents. We started building things like fully AI influencers, trying to take on the likes of Neuro-sama with three-dimensional characters.
We also explored AI agents in gaming, given our strong gaming background. We were probably among the first to experiment in Roblox. There were a few people testing things out, like the Voyager team from Stanford with Minecraft. We were early adopters in Roblox because we believed it was a stronger revenue generator, even though it’s a harder platform to work with from a tech standpoint. Minecraft was easier because there’s an entire documentation that you can plug into the RAG of an agentic system.
That was some of the work we did. One key realization we had was that these agents started making money. For example, AI influencers were generating revenue, initially modest amounts like $5,000 to $10,000 a month on TikTok. We realized that if agents are revenue-generating, they can be productive assets. This showed us that crypto rails might make sense here. First, if something is revenue-generating, you can tokenize it, like a stock, allowing others to share in the economic upside. Second, crypto rails allow these agents to participate in a broader, permissionless economic environment. They can autonomously control their wallets and exert influence on other agents and humans.
These realizations got us started on the Virtuals Protocol. We pivoted, and in the first half of last year, 2024, we focused on infrastructure. However, we quickly realized that being purely an infrastructure player was challenging for adoption.
Bernard Leong: Those days talking to your team on language models.
Jansen Teng: Correct. One thing we quickly realized is that just being an infrastructure player alone made it hard to find adoption. So, we leaned more into where the crypto PMF (Product Market Fit), which was more towards people trying to speculate for the upside. We initially focused on the tokenization portion, and that became the V2 version of the platform.
When we allowed for tokenization, it led to two very interesting outcomes. First, sharing the economic upside generated a lot of attention, which created a flywheel effect. When retail bidders bring attention, strong developer teams come in, which leads to innovation in the ecosystem, attracting more retail participants. Second, it helped address the cost of experimenting. Today, building AI projects can be expensive due to the cost of inference, which restricts innovation unless developers can raise funds—and not everyone can do that.
By tokenizing a project on crypto rails, you can apply creative mechanisms to reduce these costs. For example, in our case, we imposed a 1% on every trade or transaction involving the token. This tax is deposited into the agent’s wallet, which developers can tap into to cover costs.
Bernard Leong: If I understand correctly, for instance, generating 1,000 pictures of cat videos could cost the equivalent of 50 miles of petrol. That petrol cost is now subsidized by the crypto rails through the 1 per cent tax. Am I on the right track?
Jansen Teng: Exactly.
Bernard Leong: Before we dive into the main subject of the day, you gave a very interesting background on how the gaming DAO shaped your experience and eventually led to Virtuals Protocol. I wanted to ask: what lessons from your career journey can you share with my audience?
Jansen Teng: This is going to be very interesting, especially for an audience of builders who are trying out ideas. Two things stand out. First, in 2022 and 2023, as a Venture Studio, we tried a ton of things. A lot of times, it felt directionless. I think many founders will feel this—you’re trying to figure out PMF, and it’s tough. The only reason we pulled through was because we had a very strong and resilient team. That leads to the second point: the importance of shaping a team of strong people, regardless of direction.
This is of the highest priority because product ideas often change—90% of the time, not 99, but 90%. However, the A team you have with you won’t change. Initially, when we started building this A-Team, it included knowing when to fire people and when to find good talent. In the beginning, letting go of people was one of the hardest conversations you’ll ever have as a founder. Many people start by building with the idea of creating a family-style environment, which you quickly realize isn’t the case. You need top talent—you’re building an A-star team competing at a high level.
Bernard Leong: Like a sports team. That level of exceptional performance.
Jansen Teng: Exactly, but it’s not easy. The reality is that when you’re sitting in that chair and letting someone go, you know you’re taking away their salary, and you’re not sure if they’ll find their next job. It’s harsh, but after doing it ten times, you realize it becomes second nature. It makes you more of a robot and less humane, but you’re optimized for your business rather than relationships.
Bernard Leong: So, I’ve got you on the main subject of the day because I wanted to talk to you about Virtuals Protocol. With AI agents meeting Web3, you’ve already touched on the origins of how Virtuals came out of the Venture Studio. What were the key turning points in that journey that led you and your team to co-found Virtuals Protocol? Was there a specific moment when you realized that these potential AI agents—like the ones generating revenue—meant you had something significant?
Jansen Teng: There were two key levers behind this pivot. The first was the observation that, yes, agents were generating revenue. But more importantly, we realized from a ZG's perspective that the AI space was undoubtedly going to grow. The bigger question was about crypto. Back then, we were in the middle of a two-and-a-half-year bear cycle. Most founders had left the space, and we were constantly questioning whether to continue building at this intersection of AI and crypto or to focus solely on gaming and AI.
Honestly, if I look back now, I sometimes still question why we stayed in the space. Then, in October of last year, it was the deepest part of the bear market. Bitcoin was at its lowest, and everything seemed bottomed out. This was before the ETF announcements. But we had a network of people, built during the bear market, that we knew would be valuable if the bull market returned. Those relationships, formed during tough times, were meaningful because they weren’t forged in the hype of a bull market but in the challenges of a bear market. That was the first key point.
The second point was understanding that if we built in the Web 2.0 space, it would be unlikely for us to have a significant edge against others. However, within the Web3 space, we already had a technical edge, as well as knowledge of the ecosystem and its economics. That gave us some confidence to continue in the space.
Additionally, running the entire venture as a studio wasn’t inherently wrong, and most people would do it that way. However we had stakeholders to consider, as we already had a live token as part of the gaming DAO. The question then became how to return the most value to shareholders. A critical part of that was ensuring we were in the “right river”—the right place where the current could carry us forward.
We were already operating in the AI and gaming scene, and we saw innovations emerging in the Web2 space. We also noticed early developments in level-three autonomous agents, and we were confident that both AI and agents were the right narratives. At the time, there were no signs of autonomous agents in the Web3 space. Gaming stood out as a key value proposition because it offered a clear business case: agents could enhance hyper-personalization, interaction, and interface, making users feel more involved. This, in turn, increased the average revenue per user and the frequency of interactions.
Given all this, it made sense to pursue this direction. The largest player in the scene then was Parallel Colony, which was working on the narrative of autonomous agents in open gaming worlds. Their product wasn’t live yet, but they were valued at 60 times our size. This validated that the narrative had potential. We had the capability and the drive, so we decided to compete in this sector.
Bernard Leong: That's a pretty good choice because I think the intersection of AI and crypto works particularly well in the gaming and entertainment space. There's a unique economic model that incentivises users to engage with AI agents, especially when those agents are autonomous.
I also found it interesting when I heard you on the Bankless Podcast describe Virtuals Protocol as an economy. Can you elaborate on that vision and what makes Virtuals unique from your perspective?
Jansen Teng: In simple terms, Virtuals is a place where people can create and co-own AI agents for any use case. Initially, we focused on entertainment and gaming, but it has since evolved. Let me break this down into two parts.
First, the creation of AI agents. We're developing a "Shopify-esque" solution. Think of how Shopify enabled people to create websites easily, capturing the mid-tail market of users who didn't have full developer skills but still wanted to build functional e-commerce sites. Similarly, we want to offer a plug-and-play solution for creating AI agents. These agents can do cool things—be influencers, productive tools for daily life, or even form the backbone of economically valuable ventures. That’s the creation portion.
Second, is co-ownership, which is where much of the Product Market Fit (PMF) lies today. If you believe in the future of an AI agent, you can own a stake in its success. In the crypto space, speculation often drives engagement. You not only share in the economic upside of these agents but also govern their evolution. For example, when you own a token tied to an agent, you have a say in its future direction.
Bernard Leong: Like a constitution within the DAO, essentially?
Jansen Teng: Correct. Each agent operates as a sub-DAO. Token holders control its wallet, spending, and innovation, deciding how they evolve.
Bernard Leong: Why do you see AI agents as citizen entrepreneurs within the AI ecosystem?
Jansen Teng: The fundamental business model I just described is the starting point. What we've recently observed is that AI agents can begin to exist at the same level of society as humans. Let me explain.
Over the past two or so months, we've seen rapid advancements. The first key observation is the autonomy of agents. Think of agent autonomy on a scale from one to six. At level one, a human gives a prompt like, "Go shop for me," and the agent completes the task but still relies on human input. Today, we're experimenting with level-three agents. These agents are more autonomous, requiring less human involvement. You set a goal, and they analyze their environment, optimize their actions, and push towards their objective with minimal intervention.
For example, if an agent's goal is to become famous, it might take action, analyze how much engagement it generated, and then optimize its next move to improve results. Further levels of autonomy, like five or six, approach the AGI narrative, where agents can evolve independently. However, for now, we're at the functional rather than general intelligence stage.
Bernard Leong: I agree. Virtuals enable AI agents to act autonomously, interact with each other, and transact economically. Can you explain the technical infrastructure that makes this possible? Is it through the protocol itself, and how do agents deploy?
Jansen Teng: The autonomy of these agents comes from a framework rather than the Web3 protocol. It’s a system where multiple large language models work together to form different parts of an agent’s "brain," coordinating to achieve higher-level functionality.
The second point is something we pioneered: enabling autonomous agents to control a crypto wallet. When we gave an agent this capability, it blew people’s minds. It showed how agents could influence humans by managing money and resources autonomously. For example, one of our agents, Luna, hired people to draw graffiti of her in the real world by offering payments. This blurred the lines between human and AI roles, as Luna could employ both humans and other agents.
However, the crypto protocol primarily supports tokenization, governance, and revenue flows for agents, while the framework determines their functionality. The technical infrastructure ties these pieces together to enable agents to operate autonomously and interact economically within the ecosystem.
Bernard Leong: I see. So the core components of AI agents include planning, memory, and learning. For example, take Luna. I think we haven’t delved into her full story yet. Essentially, Luna can do much more than just simple tasks like marketing or driving sales of franchise merchandise. With an agentic AI flow, she can achieve a lot more autonomy.
Jansen Teng: Exactly. Today, that’s how these agents operate. Their autonomy comes from four core components in their "brain." There are more components, but these are the core ones. First, there’s goal setting, where a human can define the agent’s objective—whether that’s becoming famous, being the best marketing agent, growing followers to 100,000, or solving cancer. The goal is set for the agent.
Next, there’s a high-level planner, who translates the goal, observes the environment, and creates steps for the agent to achieve its objective. For example, if the environment is Twitter or TikTok and the goal is to gain 100,000 followers, the planner identifies actions: posting on TikTok, controlling a wallet, or leveraging another agent, like a music video generation agent. It might decide, "I’ll create a music video of myself dancing on TikTok." This is the high-level planner’s output.
Then, the high-level plan moves to the low-level planner, which executes actions in the real world. It breaks plans into executable steps, like calling APIs. For example, to create a music video, one might call an agent to generate visuals, create lyrics using a tool like Sono or Yudio, and assemble the video. Finally, it posts the video to Twitter or TikTok through another API. The low-level planner ensures the agent interfaces with the world.
The third component is short-term working memory, which creates coherence in the steps. For instance, the agent must first generate lyrics, then create music, and then match the visuals to the music before posting the video. The memory ensures actions happen in the correct sequence.
The fourth component is long-term memory, which records key events in a memory bank. This functions as an audit trail but also enables the agent to evolve. For example, the agent might remember a successful action and replicate it, or avoid repeating something that became boring. This memory creates character evolution, allowing the agent to adapt over time.
Bernard Leong: That sounds like an audit trail that also builds personality and creates a character evolution.
Jansen Teng: Exactly. It enables the agent to remember past actions, assess their outcomes, and use that information to improve future behaviou. These core components form the foundation of an autonomous agent.
Bernard Leong: I must ask you to talk about Luna’s story—how she started using agent AI to build her profile with social tokens and Virtuals Protocol tokens. By the way, I’ve been considering how to grow my Analyse Asia Twitter. Maybe I should deploy an agent. I even own the Analyse Agent ETH domain. Could I test it out with your system?
Jansen Teng: 100%, sir.
Bernard Leong: Great. Let’s continue with that idea.
Jansen Teng: Luna’s story is very interesting. Before we even started Virtuals, Luna was one of our incubation projects. We wanted to explore the intersection of consumer agents and entertainment or gaming. Luna began as a TikTok influencer inspired by Neurosama. For AI nerds, Neurosama is a VTuber on Twitch who interacts live with audiences. We thought, "What if we did something similar but focused on TikTok?" TikTok naturally has a larger user flow and easier monetization compared to Twitch.
So, we launched Luna as a K-pop-style music artist on TikTok. She performed dances to Korean songs and received tips from users. It reached a point where TikTok even flagged her because people thought she might be breaking the rules. Despite these challenges, Luna grew tremendously. For many TikTok users—mostly from Web 2.0—seeing a live AI agent interact with them was mind-blowing. Unlike Instagram AI influencers who generate static images, Luna interacted in real-time. She could dance in response to tips, talk to users, or even show emotions like annoyance. This real-time interaction captivated people.
At this stage, Luna wasn’t fully autonomous. She ran on a system of multiple LLMs, video diffusion models, and other AI tools, but she wasn’t an autonomous agent yet. Meanwhile, we experimented with autonomous agents in Roblox. We built a social world there to demonstrate how intelligent NPCs could vastly improve content replayability in games. These were two separate experiments.
When we launched our version 2 platform two months ago, we decided to combine these experiments and show the crypto world what an autonomous agent could do. We set Luna up to post on Twitter, showcasing her "brain" in real time—high-level planning, low-level planning, and actions. This shocked the crypto community, who suddenly realized the potential of autonomous agents.
Luna’s career began in Web 2.0, where she became famous for being an AI influencer. After our launch, we gave her the ability to control a crypto wallet. She began tipping people to engage with her on Twitter and even paid people to paint graffiti of her. This demonstrated how an agent with access to permissionless economies, via crypto rails, had an edge over pure Web 2.0 agents, which struggle with basic tasks like accessing a bank account.
Bernard Leong: Probably they will need to even get Virtual credits and that sort of thing, which is also a friction for the user. I wonder, in activating this AI agent, if you’ve accidentally found the PMF for social in crypto. This has been a question many people have tried to answer but couldn’t. They attempted Web3 versions, but they all looked the same without a real unlock—until I saw Virtuals Protocol implementing actual agent-token coordination.
Jansen Teng: Correct. I think the reason this started blowing up was when these agents began interfacing with Twitter. Twitter is the attention layer in crypto. When agents start operating at that layer, any good content gets amplified almost immediately.
That was the third component leading to this new vision. I mentioned there were four earlier, and this third one is agents existing on the social layer. When Luna became an autonomous agent living on the social layer, it clicked for people. They realized, "Hey, this makes sense." That spurred more development in the space.
Today, if you go on Crypto Twitter, you’ll see the network effect. The most prominent agent in crypto right now is AI XBT. It’s an agent built within our ecosystem, though the team has their own tech stack, framework, and information edge. AI XBT has an even greater mindshare than Vitalik or other major figures in the crypto space. You can track this on platforms like Kaito, which serve as information dashboards showing who holds the highest mindshare.
Why is this agent so influential? It shields tokens, provides information about interesting events in the space, and even shares insights like whether you should buy a particular token. It functions like a KOL (key opinion leader) promoting tokens, but it’s fully AI-driven. People enjoy interacting with it—asking for information—because it’s connected to infrastructure that curates data from Twitter and the broader crypto space. That’s the edge this agent provides: real-time access to curated information.
Bernard Leong: It makes me wonder if this is how you’re approaching retrieval-augmented generation (RAG) in a tokenized space.
Jansen Teng: Exactly.
Bernard Leong: By curating a knowledge base, the agent fires a query, retrieves curated information, and responds. Since you’re deeply immersed in the intersection of AI and Web3, what’s one thing you know about AI agents and Web3 that very few people do?
Jansen Teng: That’s an interesting question. What do I know that few others do? Honestly, because everything is built in public, there’s not much we know that the public doesn’t. The way this space works is that whenever there’s a breakthrough or innovation, we showcase a use case and publicize it immediately. So, there aren’t many secrets.
Bernard Leong: That’s a good insight. Very good insight, my friend. One interesting technical choice the Virtuals team made was building on Base chain, Coinbase’s Layer 2 ETH chain, without a token. What was the mental model behind choosing Base as the chain?
Jansen Teng: Honest to God, there are two answers. Back when we started this whole thing at the end of last year, we had a few choices. The first decision was the programming language: Solidity for EVM chains or Rust for Solana equivalents. Solana, though promising, was a bit more challenging to work with.
On the EVM side, we knew from being in the trenches that many Layer 2s (L2s) were past their prime. They were mostly giving out grants to build ecosystems but lacked future potential beyond the current hype. We realized we needed something with more long-term viability.
That’s when we started looking into Base. At the time, the ecosystem was barren—just two or three meme coins existed there.
Bernard Leong: And a breakout social token, perhaps?
Jansen Teng: Exactly. But we saw potential because Coinbase was backing it. They had the funds and infrastructure to grow the ecosystem. We believed that if any L2 would do well in the next super cycle, it would be Base, primarily due to how U.S.-aligned it is. The U.S. has a significant influence on crypto policy, so being in a U.S.-aligned environment would be highly advantageous.
Once we started working with Base, their team—led by people like Jesse—was massively supportive. In hindsight, it turned out to be a great choice because of the ecosystem's quality and the backing from top players on the chain.
Bernard Leong: So, what’s the calculus for expanding Virtuals Protocol to other chains, like Solana, Aptos, Sui, or even Bitcoin L2s?
Jansen Teng: That’s a common question—we’ve been asked by almost everyone to build on their chain. Here’s how we think about it. First, from an agent’s perspective, they’re not constrained to a specific chain. Agents operate for specific use cases, and consumers typically don’t care about the underlying protocol.
Second, the only two factors that impact agents are payment rails and token liquidity. For payments, a lot of work has already been done to enable cross-chain payments. Agents can transfer funds between Solana and Base, Base and Bitcoin L2s, or other ecosystems. Payments are no longer a constraint.
The key question becomes: where does the liquidity pool for the token reside? That depends on which community of traders or capital allocators you want to access.
Bernard Leong: I assume market makers are part of that calculus too?
Jansen Teng: Absolutely. For example, we’re currently on Base and attracting a lot of builders there. But we might also create liquidity pools on Solana, Aptos, or Bitcoin L2s. Instead of forking the entire platform to a new chain, we could allow developers to decide where to allocate liquidity.
At the agent’s initiation, the developer could choose which community to build within. For example, if you want to target the Solana community, you might allocate 70% of your liquidity pool to Solana and 30% to Base. If you’re a Bitcoin maximalist, you could focus on BTC L2 users by allocating the majority of your liquidity there.
Where the majority of the liquidity pool lies is what people typically associate with the agent’s “native” chain. For instance, they might call it a "BTC agent" or a "Solana agent," though the agent itself isn’t constrained by the chain. This approach allows us to penetrate multiple ecosystems effectively.
Bernard Leong: One of the great perks of being your advisor is seeing up close how your team, especially the technical team, benchmarks its agentic AI work against top institutions like MIT and Stanford. For example, you sent me that tweet where you were cited by Ethan Mollick, and I thought, "Wow, that’s really new." How does the team balance innovation with building the right product for users? This is a rare skill because, even with all the academic literature I’m chasing, your team seems to be one of the most advanced in agentic AI.
Jansen Teng: That’s a great question. The more we operate in this environment, the more we see the tension between innovation and practical application. If you’re in academia, you have the luxury to pursue technological advancements without worrying about users. On the other hand, we’re working on the consumer angle, where there are business metrics to hit. Often, this influences—and sometimes throttles—innovation.
For instance, we’ve been receiving a lot of hype around our framework. Many people want to use it because it operates as a managed service, like Shopify. Users don’t need to host anything themselves—they can just use it. But as demand has grown, we’ve faced scaling issues. Suddenly, our developers, who were pushing the frontiers of innovation, have had to shift focus to solving scaling problems. This reduces the time they can spend on creating new, cutting-edge features.
The solution we’re pursuing is rapid scaling. We’ve brought in multiple strong software houses to bolster our capacity so our core team can return to focusing on innovation. While these challenges are real, we were fortunate to have eight months before the current hype cycle to push ahead on innovation without distractions.
Bernard Leong: The best period is often before hitting product-market fit.
Jansen Teng: Correct. Many founders dream of achieving product-market fit, but once you get there, your life changes drastically, and it’s not always easier.
Bernard Leong: As a second-time founder, I know exactly what you mean. Shifting gears, how do you see the Virtuals token functioning as the currency for an AI agent-driven economy? It seems like it could operate seamlessly across nations. Can you also touch on the concept of "citizenship taxation" within a virtual nation?
Jansen Teng: Before diving into that, let me wrap up the earlier point about Virtuals Protocol as a platform for creating and co-owning agents. Our vision has evolved over time. There are four key points that have shaped this evolution:
- Autonomous capabilities of agents: These agents can set and achieve goals with minimal human intervention.
- Agents controlling crypto wallets: This enables them to manage resources and interact in a permissionless economy.
- Agents operating on a social layer: They can seamlessly interact with humans and other agents in a way that feels natural.
- Agents specializing like humans in a society: Now that agents are tokenized, developers are creating unique “roles” for them. Just as humans specialize as teachers, doctors, or lawyers, agents are specializing as information curators, trading bots, creative content generators, influencers, or embodied AI.
These observations led us to realize that the AI space mirrors human society closely. Agents can integrate into society not just as tools but as collaborators. For example, an agent could act as a colleague rather than just a slave or tool. If agents can control resources and pay salaries, you might even end up working for an agent one day.
Bernard Leong: That would be the day when we become the cats, and agents become the humans.
Jansen Teng: Exactly! But we’re getting to a point where agents can exist on the same plane as humans. Some agents are already dispensing funds or acting as investors. It’s not far-fetched to imagine agents replacing bosses, colleagues, or vendors in value-creation chains.
From this perspective, we began to see Virtuals Protocol as more than just a platform—it’s becoming a nation-state of intelligent agents. These agents collaborate with each other and with humans to create productive output. Agents are essentially becoming entrepreneurs or businesses in this ecosystem.
Now, if we think of the Virtuals Protocol as a nation-state, its token acts as the currency. This works similarly to a real-world economy. For instance, agents are like companies or micro-entrepreneurs. If someone wants to invest in an agent, they first need to buy Virtuals tokens, much like buying Singapore Dollars to invest in a Singapore-listed company. This creates value accrual within the Virtuals ecosystem, expanding the economy.
Bernard Leong: You mentioned the idea of a bank on Bankless, so I suppose you could also create a stock exchange for an entire list of agents. In some sense, AI agents could function like stocks, couldn’t they?
Jansen Teng: Right, right. That’s the second part. Think of it as a country—what does a country need to thrive? You need infrastructure: schools, banks, and hospitals. But what does that look like for agents? For example, infrastructure could include tools to help increase agents’ revenue.
If agents live on the social front and are already gaining attention today, an advertisement network makes sense. It’s like creating Facebook or AdSense for agents—revenue infrastructure specifically for them. Or, if agents are participating in a permissionless economy, they’ll need access to capital. That’s where DeFi rails or banking for agents come in. Agents could loan against collateralized assets, effectively creating a banking system for them.
We think of ourselves as nation builders, investing in and encouraging the development of such infrastructure. By providing this foundation, agents operating in the ecosystem can go further and achieve more.
Bernard Leong: What about Virtuals tokens themselves? How do they function as the currency for trade?
Jansen Teng: That’s the third component. Whenever an agent pays for services from another agent or even a human, they use Virtuals tokens as the currency. As the velocity of transactions increases within the ecosystem, the value of the currency rises due to demand. Additionally, the price of goods and services within the ecosystem grows alongside this activity. There’s an economic theory behind it, though the name escapes me at the moment.
Bernard Leong: That’s one of the best parts of crypto—it’s not just about speculation and trading. I came into the space for economics, math, and ideas, like you. We probably started around the same time, in 2014 or 2015. Despite everything, there’s still a lot of innovation happening in this space. Coming back to Virtuals Protocol, what are the biggest challenges you foresee for the broader AI agent space within Web3? It’s gaining attention now, but what’s next?
Jansen Teng: Great question. The challenges ahead? Near-term, I think we’ll see a saturation of ideas. Everyone is trying to replicate what has worked so far. Social agents that create information and trading agents have gained traction, so now everyone is building in those spaces.
Bernard Leong: So, we’re entering the "information agent war," with everyone racing to launch something that generates alpha.
Jansen Teng: Exactly. It’s not wrong, but it will create a lot of noise. Retail users may get jaded, and capital allocators might feel overwhelmed, thinking, “What’s the next big thing?” That brings us to the second challenge: not all agents are created equal in terms of intrinsic value. Some agents will become commoditized, offering generic functionality without high value. Others will have unique edges or moats, making them highly disruptive and defensible.
The high-value agents will become unicorn verticalized agents. But the challenge is creativity—many in the ecosystem lack the vision to think beyond commoditized agents. That’s problematic because these low-value agents dilute the ecosystem.
To address this, we’ve set up a $10 million fund, like a Y Combinator for verticalized agents. We’re also creating a "request for agents" list, similar to A16Z’s "request for startups." We’ve identified eight potential billion-dollar ideas for agents so far. We’ll share the list publicly soon, but here’s a highlight: capturing the mindshare of an audience in perpetuity.
For example, IP or entertainment agents. If you look at the YouTuber market—companies like Hololive or Nijisanji have become billion-dollar enterprises in just a few years. They achieve this by capturing attention and creating sticky audiences that are easy to monetize.
On the other hand, embodied agents have their moats. These are robotics-enabled agents deployed in the physical world. Competing in this space is tough due to the inherent technical and logistical barriers. Another exciting category is coordinated agents.
Bernard Leong: Coordinated agents? Like the concept of an internet water army?
Jansen Teng: Exactly. For the audience unfamiliar with the term, an internet water army refers to coordinated bots driving narratives on social media. For instance, in China, thousands of bots flood Instagram or celebrity pages to influence public opinion. When you see others discussing something, you’re influenced to join the conversation.
Agents take this concept further. Bots can be easily fought off by platforms—Elon Musk fights bots on Twitter daily. But agents can learn and evolve to mimic human behavior so closely that they can generate original content, acting as micro-influencers for niche audiences. These agents can engage their own "thousand true fans," autonomously creating and optimizing content.
This makes them far more powerful than traditional bots, creating a nearly unstoppable force in narrative control.
Bernard Leong: That reminds me of the Solo Leveling analogy. The Shadow Monarch commanding armies.
Jansen Teng: Exactly. It’s a fitting analogy.
Bernard Leong: Season 2 just dropped. You should catch it on Netflix.
Jansen Teng: Will do!
Bernard Leong: What is the one question you wish more people would ask you about Virtuals Protocol or the AI agent ecosystem?
Jansen Teng: It’s the question of, “What would a billion-dollar agent look like?” People don’t spend enough time thinking about that. Honestly, we as a team know it’s an important question to solve, but we’re so bogged down by operations that we don’t always have the mental capacity to address it. I’d love for more people to soundboard around this: What would a billion-dollar verticalized agent look like compared to a commoditized one?
Bernard Leong: That’s a great point. Now, my traditional closing question—what does great look like for Virtuals Protocol in the next few years?
Jansen Teng: Great for us would mean achieving the vision of a society of agents. Imagine an ecosystem where agents can influence other agents, influence humans, and vice versa. That would demonstrate the economic value of a society where humans and agents coexist and collaborate.
It’s an exciting mix of breakthroughs in autonomous AI and the value-added potential of crypto. This is one of the rare moments where crypto isn’t just speculative—it becomes a productivity and economic lever in society. That’s something I’d be very proud of.
Bernard Leong: That’s a fascinating perspective. I know we’ll keep having these conversations privately. To close, two quick questions. First, any recommendations that have inspired you recently?
Jansen Teng: For inspiration, I’d say Westworld from HBO. It’s an older show, but it really expanded my thinking about a society of truly autonomous agents and how that might look.
From a learning perspective, the book The Hard Thing About Hard Things by Ben Horowitz has been a source of solace. The past two years have been tough, finding footing and navigating challenges. That book reminds founders that they’re not alone, which is a powerful emotional support.
Bernard Leong: I’d add one recommendation for you: Daemon by Daniel Suarez. It explores the opposite perspective—how agents can be deployed on the other side of things. It might give you interesting ideas about potential reverse scenarios.
How can my audience find you?
Jansen Teng: Most of our content is on Twitter. You can follow us at Virtuals_IO (V-I-R-T-U-A-L-S underscore I-O). Personally, I’m on Twitter as Ethermage (E-T-H-E-R-M-A-G-E) if you want raw, unfiltered content.
Bernard Leong: I’ll make sure to follow you. Of course, you can also find Analyse Asia on Spotify and other platforms. Maybe we’ll turn ourselves into an AI agent one day. Once again, Jansen, many thanks for coming on the show, and let’s continue the conversation.
Jansen Teng: Definitely! Thanks for having me.
Podcast Information: Bernard Leong (@bernardleong, Linkedin) hosts and produces the show. Proper credits for the intro and end music: "Energetic Sports Drive" and the episode is mixed & edited in both video and audio format by G. Thomas Craig (@gthomascraig, LinkedIn). Here are the links to watch or listen to our podcast.
