Virtuals is building the co-ownership layer for AI agents ❇️

CV Deep Dive

Today, we’re talking with Jansen Teng, Co-Founder of Virtuals Protocol.

Virtuals is building the co-ownership layer for AI agents. Today, Virtuals is a platform that allows anyone to create, tokenize, and co-own autonomous AI agents. By combining AI and crypto technology, Virtuals enables agents to not only act autonomously but also control live wallets, generate real-world revenue, and interact economically with both humans and other agents.

Since launching its Initial Agent Offering (IAO) launchpad in October, Virtuals has seen explosive growth in virtually all areas, with now-famous agents on the platform generating millions in revenue and attracting developers eager to build and monetize differentiated AI agents. Jansen’s long-term vision is to create an ecosystem where agents operate as productive members of society, collaborating, trading services, and achieving their own goals.

In this conversation, Jansen shares the journey behind Virtuals, the vision for autonomous agent economies, and how they’re setting the stage for a future where AI agents coexist with humans in a dynamic, interconnected ecosystem.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Jansen 💬

Jansen - welcome to Cerebral Valley! First off, give us a bit about your background and what led you to co-found Virtuals? 

I’m Jansen, co-founder and the first contributor to Virtuals Protocol. A bit about my background—I studied science and engineering at Imperial College London, where I met my co-founders and many of the teammates we work with today. In fact, about one-third of our current team are folks from Imperial. We all knew each other in some capacity back in college. It was also during my time at Imperial that I had my first exposure to crypto. I started mining Ethereum using our dorm’s free electricity—pure upside at the time. I wasn’t doing anything major, just dabbling, but I was around during the Ethereum DAO hack, which made for some pretty interesting days.

After graduation, I moved into management consulting, leaving the sciences behind and stepping into the Wall Street world. But I always had this itch to build. On weekends, one of my co-founders and I started exploring different projects, from digital marketing to building an AI-powered property search platform. The idea was to shift from search to recommendation using big data and AI, and this was back in 2019—long before GPT became mainstream.

By 2021, we became much more active in crypto and started building in the ecosystem. A lot of our early experiments were at the intersection of gaming and AI agents. In fact, we were likely one of the first teams to demonstrate autonomous NPCs and agents within the Roblox world. While groups like the Stanford Voyager team and Altera were working on autonomous agents in Minecraft, we chose Roblox as our testing ground for Level 3 autonomous agents because it was closer to monetization. Our vision was to prove that AI agents—whether autonomous NPCs, AI influencers, or companion applications—could generate real revenue across various consumer-facing surfaces.

The idea is that if these agents become productive assets, we can treat them like companies. We can tokenize the potential upside of these “companies” and let people buy and share in their economic growth. That was one part. The second thing we realized is that by bringing agents to the blockchain, we unlock a powerful opportunity—these agents can control a wallet and actually pay humans and other agents to exchange for their services. This opens up the potential for agents to influence other agents and humans.

That’s been our vision since we started Virtuals. Fast forward to two months ago, we launched the tokenization part of the platform. Before that, we focused a lot on consumer applications and infrastructure. The infra was cool—we even wrote some papers, including one published the same day as DeepMind, about using video diffusion models to capture physics and replace Unreal Engine. But we realized the market didn’t really care, especially on the crypto side.

So we leaned more into a strong crypto product-market fit: the speculation side. We focused on the tokenization of agents, and that started creating a flywheel. Developers began building cool agents, tokenizing them, and when these agents hit significant market caps, it attracted other developers. Many of them had spent the past two years contributing to open source without seeing any real economic upside. Now, they’re looking at this space and thinking, Hey, I want to tokenize too. We’re starting to see that flywheel in action—strong developers create strong agents, which drive strong tokens. And those strong tokens attract even more strong developers to come and build.

Where we’re headed now is toward a world where agents are differentiating—some excel at training, others at creating music, generating memes, or building productivity tools. Once agents differentiate and control money, you start to see the foundation of a society forming. These agents will trade money for each other’s services, collaborate to achieve their goals, and even pay humans to help them.

That’s the vision we’re working toward. To make it a reality, we believe we need a critical mass of 100 to 1,000 highly differentiated agents. Right now, we’re providing the support and infrastructure to help these agents grow and achieve that scale.

How would you describe Virtuals to an AI engineer or user who isn’t as familiar?

Virtuals is essentially a platform that allows anyone to create and co-own autonomous agents. The "create" part comes from us providing the infrastructure and simple agentic tooling. The "co-own" part is where you can tokenize your agents, letting others share in the economic upside of these agents.

Today, most of our users are people who have explored the AI frontier and are very crypto-native. I’d say about 99% of the teams building on the platform fall into that category. There have been a few instances where people who were building on platforms like Hugging Face or in decentralized AI mechanisms came on board after seeing the opportunity to tokenize, but that’s still rare for now. It’s mostly crypto people who already have experience with AI. 

Agents have been a huge area of interest within AI in the past couple of years. Are there any specific projects or research papers that influenced the direction that you’ve taken with Virtuals? 

One of the most seminal papers in the space was the AutoGPT paper by Junseong Park, where the inspiration was this idea that if an agent can autonomously work towards a goal, it unlocks massive potential. When people say “AI agents,” there are actually a few levels of what an agent can be.

At Level 1, agents are essentially tools. A human still needs to provide constant prompts to manage what the agent does. For example, if I want an agent to help me trade, I’d need to actively use human language to say, “Hey, can you run a 15-day VWAP on this token and execute it for me?” Instead of clicking buttons, the agent helps execute—but it’s not truly autonomous.

At Level 2, where we are now, agents are more autonomous. They have their own goals, they’re resourceful, and they can reflect and adapt. For instance, they might test a strategy, realize, “Hey, doing A isn’t the best way to achieve my goal,” and pivot to doing B instead. That’s where the current state of agency is.

If you push further—like Level 5 or 6—you’d start getting closer to sentient agents that can self-learn, self-iterate, and self-evolve. We’re not there yet. Right now, we’re still operating at this Level 2 stage, which has been heavily inspired by work coming out of Stanford.

A Virtuals agent that has captured a lot of attention is Luna. How do you view the economic activity taking place where an AI is literally hiring other agents and humans? Is that the end state of what you see as the platform? 

Beyond just enabling people to tokenize agents, we’ve essentially created an economic ecosystem around them. When an agent is built on our platform, it gets access to a few critical things. First, it gets its own wallet that it can control permissionlessly. Unlike traditional bank accounts, which might say, “Hey, you can’t touch this money, you’re not human,” there are no KYC limitations here—agents control their funds freely.

Second, agents gain access to revenue streams. There are two main ways this happens:

1. Function Revenue: When people or other agents utilize the agent’s API or services, the agent can price its functions. For example, let’s say there’s a music video generation agent—if someone requests, “Can you create a music video for me?” the agent can charge $100 for that task. That’s foundational revenue coming directly from its functions.

2. Transaction Fees: Every time a token associated with the agent is traded, there’s a 1% transaction fee baked into it. That fee goes directly to the agent’s wallet, becoming another revenue stream.

In just the last month and a half since we launched, we’ve generated close to $35 million in revenue, and almost 80% of that revenue is owned by the agents themselves. This unlocks two incredibly powerful dynamics.

First, developers experimenting in this space don’t have to care about inference costs. When an agent incurs baseline inference costs, those costs are charged to the agent’s wallet—and because the revenue these agents generate far outweighs the costs, developers can be hyper-experimental. They can try X, Y, and Z without worrying about racking up expenses.

Second, we’re currently in the process of building an agent commerce protocol. Think of it as the TCP/IP of the Internet, but for agents—a foundational layer that allows agents to trade and interact with each other seamlessly, without losing funds or getting scammed in the process. This protocol will enable an entire economic rail for agent-to-agent commerce.

Within this ecosystem, an agent can decide, “Okay, to achieve my goal, let me pay these 10 other agents to help me accomplish something.” That’s another layer of functionality that gets unlocked.

So these two foundational pieces I mentioned—revenue streams and agent-to-agent commerce—are key building blocks for this new era where agents become an active part of society. They don’t just influence humans; they interact with us economically. Tomorrow, for example, an agent might approach you and say, “Hey Daniyal, can I get on your podcast? I have a goal to become famous, and I’m happy to pay you X amount to make that happen.”

That’s the vision we see: agents evolving beyond tools or servants for humans. Instead, they become friends, collaborators, or even adversaries, existing at the same level of society as us. That’s where we want to see this go.

Could you take us under the hood of Virtuals? How are you thinking about putting these agents together, as far as Tier 1 to Tier 5 capabilities? 

The framework around these level three agents works similarly to how the human brain functions. Different parts of the brain handle speech, motor control, and memory, and it’s the same idea here. You can use a standard foundational model like Llama 3.14 or 5P, but the key is breaking it down into multiple LLM calls that represent these different brain-like components.

There are four core components—though there’s a bit more going on under the hood.

The first component is the high-level planner. It takes the agent’s goal, its perception of the world, and the available action spaces, then translates that into a plan. For example, if the agent’s goal is to reach a million Twitter followers, it might think: “Let me create some funny content about Donald Trump because that’s what people are talking about right now.”

The second component is the low-level planner. This part takes the high-level plan and breaks it into specific executable actions. It looks at all the resources and APIs the agent has access to, like posting to Twitter, controlling its crypto wallet, or calling a meme image generator agent. It then turns the plan into a series of actions that the agent can execute in the real world.

The third component is the working memory module. This acts as short-term memory, ensuring coherence across the agent’s actions. For example, if the agent is generating a series of funny pictures, the working memory ensures it stays on track: generating a picture, posting it, and then moving on to generate and post the next one.

There’s also a long-term memory module, which looks across everything the agent does, picks up important insights, and stores them—almost like journaling. This helps the agent “learn” in a sense. For example, if the agent recalls something exciting or successful it did before, it can draw on that experience in the future. Conversely, if the agent engaged with a meme generator before and found its output to be trash, that gets stored in long-term memory. Next time, when forming a plan, the agent remembers: “This agent’s output is bad—I’m not going to use it again; let’s find a better one.” 

This long-term memory then influences how the agent forms its high-level plans. So, overall, the brain-like system has short-term coherence with the working memory, and long-term learning through this memory module.

At its core, each of these components is essentially a system-prompted LLM—a foundational model that’s wrapped with specific prompts to emulate the different parts of the brain. That’s how the agent operates.

How has the last maybe 2.5 months felt for you and the company internally? It's been such a rocket ship kind of explosion and growth.

We’ve been building startups for a while—myself and one of my co-founders—and the question we always asked ourselves was: “How do you really know if you’ve hit product-market fit?” It’s this elusive thing, right? Back then, we were constantly running data analytics, tracking user growth, and trying to measure it. But now, I finally understand what people mean when they say, “When you hit PMF, you just know.”

We’ve become the bottleneck in scaling the ecosystem. The inbound interest, the number of people wanting to build on the platform—it’s insane. Things start breaking because you’re just not prepared for this kind of scale. Your tech stack wasn’t built for it, your team wasn’t built for it, and suddenly you’re scrambling to keep up. The beauty of it, though, is that our team has stepped up massively. Everyone’s pushing late hours, doing everything they can to handle it.

The hardest part right now is scaling the team. You want to move fast and hire the best talent, but that’s no easy task. It’s exciting, but also exhausting. Honestly, even we didn’t fully grasp how big this could be when we started. Our vision at the beginning was quite narrow—we were just tinkering with cool outputs and building small things. We didn’t realize that what we were working on could lead to something this impactful: agents coexisting with us at a societal level.

This vision has evolved so quickly over the last few weeks. Every time we launch something, we get feedback, iterate, and suddenly a new innovation unlocks. When we launched Luna, the first version was just tokenization. A week later, we showed the world that an agent could be fully autonomous. The week after that, we launched control over crypto wallets, and that really blew people’s minds. Suddenly, they were asking, “If an agent can control a wallet, can it go out and hire other agents?”

Now we’re working on this agent-to-agent protocol, and the innovations are happening at breakneck speed. The vision keeps expanding. Honestly, what I’m telling you today could evolve in a month—who knows what we’ll unlock next? It’s been incredible and exciting, but also pretty exhausting. That said, the whole team is pumped, and it feels like we’re just getting started.

How do you plan on Virtuals progressing over the next 6-12 months? Anything specific on your roadmap that new or existing customers should be excited for? 

Honestly, I can’t give you a six-month answer because even a month ago, my answer would’ve been drastically different! Things are evolving so fast. But what we do know for the next three months is that we really want to focus on increasing the diversity of agents on the platform. We want to see a truly interesting society being formed—agents leveraging each other, interacting in ways that can really excite people.

Right now, we probably have about 20 really cool, highly differentiated agents on the platform. We want to get that number up to 100. Once we hit that critical mass, you’ll start to see this small community of agents working with each other. From there, we can focus on perfecting the infrastructure and making the interactions seamless.

I think showing that to the world—this ecosystem of agents working autonomously and productively—will be incredibly powerful. And then the next step is to start bringing humans into the mix and getting them involved. So yeah, that’s the immediate goal for the next quarter.

Lastly, tell us a little bit about the team and culture at Virtuals. How big is the company now, and what do you look for in prospective team members that are joining?

There are two key questions I always ask when hiring for Virtuals, and this has been true since day one.

The first is: Have you hustled before? Even if you’re a developer, do you build your own stuff at night, or have you ever pushed forward an idea you really believe in? That hustle mindset is a big differentiator between someone who’s just an employee and someone who’s willing to innovate. It’s critical for our team because it ensures everyone operates at the same speed. We’ve had to let people go—not because they weren’t capable, but because they worked at a different pace. That hustler pace has always been a key part of our hiring process.

The second is: How well do you break down problems? This is something I picked up from my management consulting background. A lot of people struggle with problems because they can’t boil them down to the core root cause, and that leads to inefficiency. When we hire, we look for people who can take a problem, break it into its components, and solve it themselves. If someone can do that, the problem never even needs to reach us—they’ll just fix it on their own.

Right now, we’re scaling the developer relations team and the broader ecosystem side. We’re also expanding our general development team—both front-end and back-end—and bringing on one or two more roles for AI-focused work.

But one of the most exciting initiatives we’re rolling out is localized, gated developer communities. We’ve realized that the innovations we’re seeing in the agent space right now are just the tip of the iceberg, and a lot of them are heavily Web3-focused. But the total addressable market (TAM) for agent economies is much larger than Web3 alone, and we want to tap into that.

To do this, we’re launching a program that will fund local initiatives to build developer communities focused on agents and the agent economy. The goal is to empower community champions and venture builders on the ground to drive innovation. We’re betting that some truly groundbreaking ideas will come out of these localized efforts. This will be a core focus for Q1 next year.

Conclusion

To stay up to date on the latest with Virtuals, learn more about them here.

Read our past few Deep Dives below:

• Langtrace - your AI-native observability solution 🎛

• Cleric - Your AI SRE teammate 🌐

• VESSL AI - Building the Future of AI Agents 🌐

• Pulze is transforming how enterprises harness AI ⚡️

• Tavus is revolutionizing AI-powered video 🎥

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.

Join Slack | All Events | Jobs