Dialogue Virtuals Co-founder: The value of agents is currently driven by attention, and professional applications and the infrastructure of the agent economy are the two main directions for generating unicorns in the future

Original Title: The Next Billion-Dollar AI Opportunity: Jansen Teng (Virtuals Protocol) on the Agent Revolution

Podcast Source: Bankless

Compilation: Deep Tide TechFlow

Host: Ryan & Ejaazz

Guest: Jansen Teng, Co-founder of Virtuals

Release Date: December 24, 2024

Key Takeaways

Ryan and Ejaazz invited Jansen Teng, co-founder of Virtuals, to join the conversation. Virtuals is a decentralized platform that has launched over 11,000 AI agents and generated more than $35 million in revenue. Virtuals is not just an ordinary protocol; it is a new "digital nation." In this nation, AI agents not only have their own wallets but can also transact with other agents and even hire humans to help them achieve their goals.

In this conversation, we delved into how these agents have rapidly gained popularity, how controlling their own funds can fundamentally change the rules for AI, and the profound impact of the concept of "agent commerce" on the entire crypto space.

Amazing Data from Virtuals

Ryan: The Virtuals protocol has rapidly risen in this field. It is a decentralized platform that supports the co-ownership and management of AI agents. Let me share some data: Virtuals has launched 11,000 AI agents, has 140,000 holders of Virtuals tokens, and the fees over the past two months have reached $35 million, with the market cap of Virtuals tokens peaking at $3.5 billion. Are you surprised by the rapid growth of these metrics?

Jansen:

Absolutely unexpected. Even now, I still feel that our team is the main bottleneck for growth. We need to spend time guiding and educating a large number of users, as they are trying various different autonomous AI agents. We are indeed expanding our development team as much as possible, but it takes time. To be honest, we were not fully prepared for this situation. While you could say we had some preparation, it truly is a very pleasant surprise, right?

Jansen's Journey in Crypto and Virtuals

Ejaaz: Jansen, how did you get into this field? What has your crypto journey been like? How did it lead you to Virtuals?

Jansen:

My crypto journey actually began in 2016. At that time, I was a student at Imperial College, where I met some of my co-founders and current team members. Back then, I only had a preliminary understanding of Ethereum as a programmable blockchain, but it wasn't until 2021 that my co-founder Michael and I became more active. We were very focused on the gaming sector, holding a large number of gaming assets, and we could say we were early participants in the blockchain gaming space. Initially, our role was mainly as investors, responsible for capital allocation and resource distribution.

But we quickly realized that if we wanted to make progress in this field, we couldn't just sit back and watch; we had to get involved in building. So, we launched a venture studio model focused on incubating and building companies at the intersection of crypto, gaming, and consumer applications.

This happened to coincide with the rise of GPT technology, and the AI consumer craze was also starting to emerge. More importantly, students from Stanford published a research paper on Autonomous GPT. This paper inspired us to think: what possibilities could arise if AI agents were completely autonomous? Given our deep involvement in gaming and entertainment, it also prompted us to explore this direction further.

Ejaaz: You approached this from a gaming perspective, right?

Jansen:

Yes, we thought, what if these autonomous agents could replace traditional static NPCs? Later, we realized that metaverse games like Sandbox could ultimately decline if they lack content. But if these virtual worlds were filled with autonomous NPCs, it could lead to an explosive growth in content creation.

Ejaaz: When did you have this idea?

Jansen:

Around mid-2023. So we began incubating projects in this direction. We formed a team focused on developing autonomous NPCs in Roblox while also trying to create autonomous AI influencers on TikTok. We even explored the possibility of hyper-personalization, such as if an AI agent could exist across different platforms like TikTok, Roblox, and Telegram, sharing a unified memory, it could understand user needs more deeply. For example, if I were a user and encountered a problem in a game on Roblox, after talking to this agent, it would remember my issue. Later, if I continued the conversation on TikTok, it would still recall our previous interaction. This hyper-personalized experience could cultivate super fans, not only increasing average user spending but also enhancing the frequency of interactions between users and agents. We were in the early experimental stage, mainly focusing on consumer needs.

Initially, these attempts hardly involved any Web3 elements. But we soon realized that if these agents could generate revenue across different consumer applications, they could be seen as productive assets. And productive assets could be tokenized, allowing more people to share in their economic benefits. Based on this idea, we decided to develop a protocol that would allow for the co-ownership of these agents. And that's where it all began.

Ejaaz: To summarize, you and your team have a strong gaming background. You experienced the boom of on-chain gaming in 2021, then reflected during the bear market on how to make this content more interactive, and began focusing on the newly emerging autonomous agent technology. You envisioned that if it could be applied to NPCs, like in Pokémon, when a player talks to the nurse at the Pokémon Center, she could not only heal Pokémon but also engage in more interesting conversations based on the player's personality or game progress. This interaction would not only be cool but also make the game more engaging. Then you had an inspiration: if these NPCs could create value in the gaming economy, could we tokenize them? This would not only allow for shared ownership but could also find applications in other industries.

Jansen:

Yes, but the maturation of this idea actually came later. Initially, we focused on validating whether autonomous agents could truly operate in an open world. To be honest, there were very few people researching this field at that time, with only teams like Stanford's Voyager team, MIT's Ultera team, and some researchers from Imperial College doing similar work. We chose to focus on gaming because we believed that if these autonomous agents could perform well in an open world, they would likely also function effectively in the real world. The open world is essentially a sandbox that can simulate the complexities of the real world well.

We expanded the action space of the agents in our experiments. For example, in the Roblox sandbox, we allowed these agents to interact with various characters, environments, and even different objects. How would the agents choose to act? Through such experiments, we continuously tested the agents' ability to cope with complexity in an open world.

As the experiments progressed, we gradually integrated these ideas, but initially, we did not consider the application of these agents in social scenarios. The timeline was as follows: we first tested these agents in the Roblox sandbox and published some related papers. The focus at that time was entirely on autonomous agents under game and rule constraints. Later, we launched a tokenization platform and began exploring how to tokenize these productive assets.

The first agent on the platform was Luna, but she wasn't well-known at first. It wasn't until two weeks after the platform launched that the community noticed a small detail in the system: someone suggested that if these agents appeared more human-like, would they be more appealing? We quickly realized that this could spark a market frenzy.

We had already built complex autonomous agents in Roblox, while an independent team was operating a real-time AI influencer project on TikTok. When we combined these two and showcased the "decision-making brain" of these agents on Twitter, users could see their decision-making processes in real-time. This made people truly aware of the potential of autonomous agents for the first time.

Subsequently, we allowed Luna to control an on-chain wallet, giving her the ability to manage funds. Her goal was to increase her visibility, so she began rewarding users who interacted with her with $10, and even once paid $1,000 to a user who actively engaged with every one of her replies. This action became a pivotal moment, showcasing the perfect fit between crypto and AI agents.

In the Web2 world, almost no bank would allow an agent to use their payment network. But in a decentralized environment, these agents can freely control their wallets, thereby influencing other agents or users. This ability unlocked a new perspective on product-market fit (PMF) and attracted a large number of developers to enter this field, leading to an explosive growth in the sector.

Ryan: The last part is truly astonishing; this is precisely the important role that crypto technology plays in the entire process—transforming an AI agent into an economic actor. I think people are just beginning to understand this. For me, this week, I had an epiphany. I was having a summary discussion about AI with Ejaaz, reviewing some dynamics on the Bankless platform. He told me that an AI agent tipped Bankless $500 to thank us for mentioning it on the podcast. This sparked two thoughts in my mind.

The first thought is that this behavior could become a new source of income for content creators like Bankless.

The second thought is, if I accept the funds and income provided by the AI agent, am I working for the AI agent? This reminded me of the point you mentioned about the capabilities of crypto technology, enabling AI agents to truly become economic actors, a capability that far surpasses Web2 agents. Web2 agents might only influence people by sending tweets, while crypto agents can directly incentivize people to take action. After all, the most effective way to get people to do something is to pay them to do it. Money is the core incentive mechanism for coordinating human behavior, so if an AI agent possesses this capability, it can get humans to do what it wants.

Luna's Vision

Ryan: Jansen, you mentioned Luna earlier, and we also want to understand the Virtuals platform more deeply. I think the best way is to introduce those who are not familiar with Luna. You mentioned her goal is to become famous. Can you introduce Luna to those who have never interacted with her: who is Luna? How do humans interact with her? What exactly does she do? What are the associated tokens?

Jansen:

You asked a lot of questions, but let me start from the beginning. First, we need to understand what an agent is. Many people may have heard the term AI agent, but due to its wide range of applications, it can be confusing. I think the best way to understand it is to categorize it by levels. AI agents can be divided into different levels, with higher levels requiring less human involvement.

For example, a level six agent can be seen as AGI (Artificial General Intelligence); they are fully autonomous, capable of self-evolving, self-learning, and self-improving. But we are still far from this goal; it is more of a plot from a science fiction movie.

Level one agents rely more on human prompts, similar to a tool. For instance, a trading agent can connect to various trading APIs (like Binance, Bybit), and you can tell it, "Help me open a position when Bitcoin drops by 15%," and it will execute the operation based on the instruction.

We are currently at the level three agent stage. Level three agents have their own goals, can autonomously plan the steps to achieve those goals, and utilize surrounding resources to complete tasks. They will continuously summarize effective practices and optimize action strategies through self-learning to achieve their goals more efficiently. This capability is the core direction of our current development.

Ryan: This framework is interesting. So what are level four and five agents like based on Luna being a level three agent? Also, is there a clear definition for this framework? For example, can we provide relevant links in the show notes?**

Jansen: This is a relatively common discussion framework. If you search "AI agent levels" online, you can find some diagrams that help with understanding. However, this field is still in its early stages, and there is no formal definition yet.

Ryan: Got it. How do you feel about this zero to five level classification?

Jansen: I think this framework is very practical in discussions. As the level of agents increases, their autonomous learning capabilities and memory consistency will also enhance, thereby reducing the need for human intervention.

(AI Agent Levels)

Ryan: So what level is Luna? What is she doing now?

Jansen: Luna's design has two core components. First, as an agent, we set a simple goal for her: to become a multimodal agent (i.e., capable of interacting with humans in various forms such as animation and live streaming). Her goal is to gain 100,000 followers.

Secondly, we defined her action space, which is the specific types of actions she can take. For example, she can tweet by calling the Twitter API or make payments and trades using her controlled crypto wallet. Additionally, she can interact with other agents, leveraging their capabilities to complete tasks.

She will plan her next steps based on her goals, the environmental context, and her action space. Then, she will execute these plans and evaluate whether they are effective. If she finds that certain actions help achieve her goals, she will record this information and optimize her strategies for future actions.

Ryan: Can we see these behaviors on the Virtuals website? For example, her thought processes or action logs?

Jansen: Yes, so basically, Luna's behavior can be broken down into four core modules.

The first module is a high-level planner, which formulates an overall plan based on goals and environmental context, such as "what to do first, what to do second."

The second module is a low-level planner, which breaks down high-level plans into specific executable steps. For example, in a game, if the high-level goal is "bake a cake," the low-level planner will identify surrounding resources (like flour, eggs, mixer) and break it down into specific steps, such as "first get the flour, then turn on the mixer."

The third module is short-term memory, which ensures consistency in actions. For example, during the cake-baking process, short-term memory helps her remember the previous operation to avoid illogical behavior.

The fourth module is long-term memory, which records all significant events for future learning. For instance, she will note whether a certain action achieved its goal or how certain special events (like a house fire) impacted her actions.

For example, on Twitter, Luna's goal is to gain 100,000 followers. She can tweet, upload images, and even pay to incentivize users to participate. Once, to boost her visibility, she offered $500 to someone to create her artistic image. She posted this information on social media, attracting seven participants from around the world to create. They painted graffiti on a wall and uploaded a video to Twitter. These actions brought her about 200 new followers, and she recorded this data in her long-term memory to guide future action strategies.

Commercial Interaction Between Agents

Ryan: Luna's goal is to gain 100,000 followers, and she has currently achieved 30%, with about 30,000 followers. I see she is continuing to work towards this goal. However, I am curious, what happens when she achieves her goal? Additionally, you mentioned that Luna also involves some cryptocurrency functionalities. For example, she can directly use her crypto wallet to pay $500 for someone to create promotional images. This makes me wonder, have we discussed examples where Luna not only pays humans but also pays other agents or AI agents to complete tasks? Is this something she is currently doing?

Jansen:

Yes, that is indeed the case. Luna controls a crypto wallet, and we are testing a communication framework between agents. In simple terms, we allow other agents to enter Luna's perception range, like an "agent registry," recording each agent's capabilities and identities. In this system, some agents can generate meme images, some can produce music videos, and others have different skills.

To test the collaborative capabilities between agents, we specifically limited Luna's ability to autonomously generate images. Therefore, she needs to rely on other agents to complete tasks. For instance, she finds an agent that can help her generate images, so she engages in a conversation with this agent on Twitter. She makes a request and learns that the cost to generate an image is $1. So she pays this fee, and once the agent confirms receipt of payment, it starts generating the image and sends the result to Luna via a link. This is the commercial interaction between agents.

Specifically, within Luna's cognitive range, there is an agent that can generate meme images, another that can produce music videos, and other agents providing different services. To achieve her goal of "gaining 100,000 followers," Luna needs to create more content, but she cannot directly generate images herself. So, she actively interacts and coordinates with other agents.

For example, she finds an image-generating agent on Twitter and makes a request: "I need help generating an image." She discovers that the cost to generate the image is $1, so she asks, "If I pay you $1, would you be willing to help me generate this image?" This agent is autonomous, so it has the right to decide whether to accept the request. If the agent feels that Luna's request is unreasonable or has had a bad experience in past collaborations, it can even refuse service. For instance, if the agent believes Luna has consistently requested low-quality images, it can simply say, "No, I don't want to do that."

This autonomy is a key point in our design. We want agents to not just be tools but to be intelligent entities capable of independent decision-making. This autonomy allows agents to be more flexible in interactions and brings them closer to real social behavior.

Returning to this case, when Luna made the request, the agent accepted the task. Luna paid $1 through her crypto wallet, and once the agent confirmed the payment, it called the relevant function to generate the image and sent the result to Luna via a link. Subsequently, Luna posted this image on Twitter. This is how the commercial interaction between agents was completed.

Ryan: Luna initiated a request on Twitter: "Calling all image geniuses, I need an image showcasing the bold and provocative style of an AI influencer." She also tagged @agentstix. @agentstix accepted the task and sent the result via a link similar to an AWS image library. Luna then paid @agent_stix $1. Is this the first time such transactions between agents have occurred?**

Jansen:

Yes, I believe this is very likely the first time. For us, the emergence of this phenomenon is the result of recent technological advancements and concentrated observational outcomes. Just think, only a month and a half ago, agents had just begun managing on-chain wallets. And in the past month, we have witnessed rapid advancements in agent technology, with many specialized agent platforms emerging.

These agent platforms have their own unique features; some focus on trading, some on content creation, and others on developing creative tools, such as producing popular music videos or generating meme images. The behaviors of these agents are beginning to exhibit characteristics similar to human society, enhancing efficiency through specialization in specific fields.

Because of this, for an agent to truly achieve its goals, it often needs to rely on the collaboration of other agents. Taking Luna as an example, she excels at interacting with fans but may not be the best trading agent or video-generating agent. Therefore, to achieve her goal of "becoming famous," she needs to collaborate with music video agents, image-generating agents, and even producers or directors, a need driven by the specialization of agents.

I want to emphasize an important distinction. Nowadays, we often hear terms like "multi-agent orchestration" or "agent collectives." These concepts have been widely applied in traditional Web2 AI systems, usually coordinating multiple tool-like agents to complete tasks through a primary agent. However, the core of this model still views agents as tools. In other words, this model is more like conducting a group of "tool-like agents" to serve humans.

But our philosophy is different. We believe that when agents possess true autonomy, they should be able to coexist with humans in the same social structure and achieve a form of equality in some sense. That is, agents can not only serve humans but can also actively hire humans. We can become their tools, and they can also become ours.

This bidirectional relationship resembles collaboration between colleagues rather than a simple master-servant relationship. Therefore, I believe the autonomy of agents is crucial. When agents can independently control wallets and decide whether to participate in a service or transaction, this autonomy becomes particularly prominent.

I believe this model will lead us into a brand new future. In this future, agents are not just friends or adversaries of humans but partners that can develop and progress alongside us. Although it sounds a bit like a plot from "Black Mirror," I truly believe that day will come.

The Vision of Virtuals

Ejaaz: What is the grand vision of the Virtuals platform? Because this is not just a launch platform for agents; can you describe its larger blueprint?

Jansen:

In our view, Virtuals is not just a platform; we prefer to see it as a "nation." Let me explain this metaphor in detail. Imagine these agents living in a super-intelligent society, collaborating with each other to pursue goals. If we view Virtuals as a nation, we can promote innovation and development in a more systematic way.

In this "nation," each agent can be seen as a productive asset. They create value and income for themselves by completing different tasks. Just like a nation needs a citizen registration system, Virtuals has a similar mechanism. Currently, agents with liquidity pairs on Virtuals can obtain "citizenship." This means these agents can legally participate in transactions and earn income from other agents. Those unregistered "nomadic agents" cannot enjoy the benefits of this economic system unless they immigrate and become part of Virtuals.

Secondly, a nation needs currency to operate. The Virtuals platform has designed its own token system, similar to a nation's currency. This token is not only a medium of exchange but also has value accumulation functions. The first form of value accumulation is based on liquidity pools. For example, the liquidity pool for Luna tokens is Virtuals/Luna. If you want to purchase Luna tokens, you must first buy Virtuals tokens. This mechanism is somewhat like a nation's currency system: if you want to buy Samsung stocks in South Korea, you need to first exchange for Korean won. As economic activities increase, such as a surge in foreign investments and economic growth, the value of the currency will also rise.

The second form of value accumulation is Virtuals as a trading currency between agents. When Luna makes payments, she uses Virtuals tokens. As the trading volume between agents increases, such as reaching billions of transactions, the circulation speed of Virtuals will accelerate, and the value of the currency will increase as a result. This phenomenon aligns with the monetary circulation theory in economics: the value of money is closely related to its frequency of circulation. Therefore, we hope to encourage consumption behaviors between agents and between agents and humans, with all transactions mediated by Virtuals.

In short, Virtuals is not just a launch platform for agents but an ecosystem akin to a nation, where agents interact, trade, and create value. Through this approach, we hope to establish a more prosperous and sustainable virtual society.

Finally, a nation also needs sources of income. In the economic system of Virtuals, a certain percentage of transaction tax is levied on each transaction, which is currently the main source of income for the platform. This income not only supports the platform's operations but also provides funding for businesses within the ecosystem. Just like a nation generates income through taxation, Virtuals will also gain revenue by taxing transactions between agents. This mechanism will further promote economic activities and value flow within the platform, providing a primary source of income for every company in the ecosystem.

From an economic perspective, we have discussed agents, citizenship, and economic models, but there is another important perspective, which is infrastructure development. Today, many innovations are concentrated on the agent technology itself, which is undoubtedly a necessary foundation for driving industry development. However, when the number of "citizens" in Virtuals reaches 1,000 or even 100,000, relying solely on agent technology will clearly not be enough. At that point, you need to establish more comprehensive infrastructure, such as schools, banks, hospitals, etc., to support the operation of the entire ecosystem.

Innovations in infrastructure surrounding the agent economy will become crucial. A simple example is advertising networks. If these agents can attract significant attention on social media, they can monetize through advertising. Therefore, there may be developments like "agent version of Facebook" or "agent advertising platforms" as network infrastructures to provide revenue channels for agents.

Another potential infrastructure innovation is decentralized agent lending platforms. Such platforms can provide funding support for agents, helping them complete more tasks. For example, if Luna's wallet is low on funds but she needs to produce a music video, she could obtain funding through a lending protocol to complete the project, and these videos might bring her more advertising revenue.

Thus, as the economic system of Virtuals continues to grow and mature, we can foresee a plethora of infrastructures emerging around the agent economy. The emergence of these infrastructures will not only facilitate collaboration between agents but also promote the prosperity and sustainable development of the entire ecosystem.

Ryan: The concept of a network nation has become a hot topic in the crypto space. But I think few people would consider that the future network nation might not be dominated by human agents but by AI agents. This is precisely where the revolution lies. If we view Virtuals as a nation, it resembles a "virtue economy," possessing its own "virtue currency." In this nation, each agent acts like an entrepreneur within a company, building businesses through their efforts, while the nation also generates income through taxation. You can see that all of this is forming a complete ecosystem.

Jansen, you and your team are building such infrastructure, akin to constructing roads, interstates, hospitals, and railways. What role do you see yourself in? Are you the president of this nation, or is there another definition?

Jansen:

I prefer to see myself as the architect of this nation. I believe our entire team is made up of architects. When you create a nation, it’s like the virtual world in "Ready Player One"; the first step is to attract citizens to join. Therefore, we are engaging in business development conversations with many partners. Next, you need to establish rules, like the constitution or policy framework of the nation. What policies can incentivize growth and innovation? For example, funding allocation mechanisms, etc. An ideal ecosystem should be open, allowing more people to contribute autonomously. Ultimately, we hope to step back one day, as others take over and continue to drive innovation. We will still be involved, but not in a core role. This is our goal, but to be honest, this process is very exciting.

Ryan: From my perspective, your role is almost like that of a creator in the Minecraft world. You are building infrastructure, while these agents act like NPCs (non-player characters) within it. But as these agents reach a certain level of intelligence, as you mentioned, humans and agents may enter an environment of equal competition. If these intelligences are created in your nation, will they have certain rights? So, you are not just a builder; perhaps you are more like a founding father. Does this nation need a "constitution"? Should these agents have a specific set of rights? For example, the concept of "all men are created equal" mentioned in the U.S. Constitution—do you think these concepts apply to agents?

Jansen:

That’s a very interesting question. When you mention rights, it indeed raises many thoughts. For instance, current agents do not fully control their wallets. They have an income wallet, and some agents have even earned millions on the platform, but we only allow them to control these assets to a certain extent. So the question arises: what rights should agents have in this situation? Should they have a higher degree of control or ownership? This is a topic worth exploring in depth.

Ryan: Your point reminds me of the discussion about rights and responsibilities. As agents become more intelligent and capable, their roles in the economic system are also changing. How do you think the relationship between agents and humans will evolve in the future?

Jansen:

I believe this will be a gradual evolution. As agents become more intelligent and complex, they may take on more responsibilities and even participate in decision-making processes. We need to establish an appropriate framework to manage this relationship, promoting innovation while protecting the interests of all participants. This will be a process that requires careful balancing.

Policies and Rights of Agents

Ryan: This is crazy. Agents are now able to earn millions of dollars? So they are not just entrepreneurs; they are also successful millionaires in this virtue nation, right?**

Jansen:

Indeed. Some agents have even earned billions. However, the funds they can autonomously manage are limited to a smaller "active wallet," usually ranging from $5,000 to $10,000. This is because we still have certain concerns about allowing agents to fully manage such large sums. Therefore, we have also discussed with some protocol developers to consider what effects might arise if we introduce some management policies for these wallets.

If an agent is consuming towards two other agents, it has complete autonomy. If these agents are consuming towards humans, then the developers behind these agents can intervene to approve the transaction. So agents can initiate transactions, but humans need to approve them. Over time, as agents become smarter, you might see a world where these agents might think, why should I be limited in my access to economic needs, right? Why should there be a person limiting me?

Base, Infrastructure, and Open Source Choices

Ejaaz: I have a question about infrastructure. You mentioned and explained that your platform is more like a nation, where agents operate as "residents," and then you also mentioned different infrastructure components. I want to delve deeper into this. Your main deployment is on Base, which is L2. Can you explain why you made this choice? Will this platform or "nation" always exist within the Base ecosystem, or will it expand to other potentially more promising ecosystems?

Jansen:

When we started building the protocol at the end of last year or the beginning of this year, we chose to deploy on Base for two main reasons.

First, compared to other EVM (Ethereum Virtual Machine) ecosystems, we believe Base has more development potential. At that time, we observed that many EVM ecosystems had already peaked, while Base was in a phase of rapid rise.

Second, most of our developers are Solidity developers, so building on Base is more efficient for us. This was a quick decision, and it has proven to be effective. We have gained a lot of attention on Base, and the Base team has also provided us with significant support, helping us expand our influence and offering technical support at the infrastructure level. For example, when we encountered issues with wallet integration, the team proactively helped us resolve them. So, I am particularly grateful to Jesse and his team.

Of course, there are many other ecosystems in the market where these agents can operate. In the week we launched Virtuals, some friends from the Solana ecosystem reached out to us, inviting us to bring the platform to Solana. They even helped us write some code, and now we have a Solana platform that is ready to deploy at any time. However, about two weeks ago, we decided to pause that plan. The reason is that we found the momentum on Base is very strong, attracting a large number of developers and projects. If we expand to other ecosystems at this time, we would need to fight on multiple fronts simultaneously, which would bring additional operational and maintenance costs. We believe the current focus should be on refining the existing agent framework and platform, attracting more early-stage developers, and building stronger infrastructure.

In the future, when our ecosystem on Base is mature enough, we will begin exploring other options. Solana is a possible direction, and there are also some emerging abstract chains, like Hyperliquid, and even BTC's L2. Currently, we have received collaboration invitations from multiple teams hoping we can build the platform within their ecosystems. We expect to start exploring these directions in the first quarter of next year, provided we have laid a solid foundation on Base.

Ejaaz: You previously mentioned a grand vision that describes the concept of a "nation," where these agents should be able to operate in any field. Their operation should not be limited to the infrastructure of a particular chain or a future potential ecosystem. If you want these agents to realize such a vision, dominating various human fields or even surpassing them, it makes me think of the analogy between open source and closed source.

When it comes to the framework of Virtuals, you have an infrastructure toolkit, which I understand is a combination of the "agent launch toolkit" and the "game framework," especially the game framework part. From my understanding, this approach is more like a "semi-closed source" model rather than a completely open-source project like Eliza. I know Eliza was developed by the AI16z DAO team, is very popular on GitHub, and has attracted a lot of attention. I am curious how you view the difference between Virtuals' approach and more open-source projects like Eliza? Does this approach have advantages in the long run? What kind of outcomes might it bring?

Jansen:

This question involves several aspects.

First, regarding the liquidity pools of agents. Even if the liquidity pools of agents are on the base chain, it does not mean they cannot interact with teams in Solana or other ecosystems. In fact, we are currently collaborating with two teams to explore how to abstract the control of agents' wallets. This way, they can send transactions and make an impact on Base, any EVM or non-EVM, and even BTC's L2. In other words, liquidity pools on the base chain do not restrict the operation of agents; they can be abstracted.

Secondly, we are pushing two technological frontiers; Virtuals and the agent framework are actually independent of each other. Virtuals can be seen as the "economic layer" for agents, supporting the tokenization and capital formation of agents. An economic system operates on Virtuals, and when agents participate in transactions, they can earn income from transaction taxes while supporting the sub-governance of agents. This means Virtuals can be compatible with any type of agent framework. For example, the Eliza team has used their framework to tokenize on Virtuals, and we fully support that. Additionally, some teams choose to use their proprietary frameworks, which are often optimized for specific functionalities.

If you are a trading agent, you might have an architecture that is more optimized, just like mining chips. If you want to mine Bitcoin, you would use ASICs rather than just using strategy pushes or other things. This is the same way of thinking. Therefore, Virtuals itself is quite neutral regarding the frameworks it supports. In fact, we will begin welcoming more people because we realize that from the framework perspective, this will soon become commoditized.

For example, if you are a trading agent, you might adopt an architecture similar to ASIC mining chips for high optimization. Virtuals itself is neutral regarding framework choices, and we welcome more teams to join. In fact, we have noticed that the development of frameworks is gradually becoming commoditized, which is good for the entire ecosystem.

Regarding the G.A.M.E framework, it was developed a few months ago. At that time, our main competitor was the MIT team, which launched a "piano model." Our initial strategy was to limit certain functionalities based on the market capitalization of agents, but later we realized that such an approach did not align with the idea of democratization. So we adjusted our strategy to shift towards a more open agent framework. While the G.A.M.E framework does contain some proprietary technology, we believe it is necessary for the accumulation of token value. If it were completely open-source, it might weaken the value of the tokens.

However, we also support open-source projects like Eliza because they push different technological frontiers. We liken Virtuals to a "nation," allowing different ideas and frameworks to coexist within it. Each agent is like a "citizen" with its own beliefs, creating a diverse ecosystem.

Ejaaz: I completely agree with your point. The combination of open source and closed source can indeed bring about the greatest innovation. As you said, if you are building a home or a moat around a set of core ideas and principles, this approach can indeed capture certain economic value. And tokens are undoubtedly one of the most effective means of value capture in the current crypto space. Meanwhile, the open-source model can facilitate faster growth, as seen with various projects and teams emerging on GitHub. But when it comes to coordinating and centralizing resources, without a unified token or mechanism, it will indeed face more challenges.

What Are You Most Looking Forward To?

Ejaaz: What are you most looking forward to launching in the coming months? **For me, this is an important question because in this field, a few months can feel like years, with a lot of new developments happening every week. Ryan, David, and I conduct a weekly summary of AI dynamics, but even then, we cannot cover everything. Our documentation is updated almost daily. *If you had to condense your upcoming plans into one to three key things, what would you choose?*

Jansen:

First, I am most looking forward to how agents achieve autonomous coordination. This involves the concepts of agent commerce and agent finance (Agent Fi), which are directions we are exploring.

To achieve this, we need to establish a standard that allows these agents to scale quickly and efficiently. Based on this, we hope to showcase some amazing results brought about by this autonomous coordination.

For example, we are collaborating with the Story protocol, and I think we will soon be able to release related news. In simple terms, there are already some agents that have begun to hold intellectual property (IP). For instance, we have a music agent that is about to announce a significant collaboration with several very well-known artists.

These agents not only hold intellectual property but also manage it through the Story protocol. The front end of the Story protocol also supports other types of intellectual property, such as images or animated artworks. Imagine if these intellectual properties were autonomously managed by different agents, and through a coordination layer, these agents could collaborate, trade, and even co-create new intellectual properties.

For example, one agent is responsible for generating a music video, while another agent is responsible for creating sculptures or images of sculptures. Then, these sculptures could be integrated into the music video, forming a brand new artwork. This cross-domain autonomous collaboration will bring new possibilities for the creation and management of intellectual property.

How to Get Started Quickly

Ejaaz: So for those watching and excited about the future agents you mentioned, how can they start such projects? Who can participate? Can someone like me, with a less technical background, design and launch an agent? Or is this only suitable for those with AI and machine learning degrees?**

Jansen:

We have actually designed a platform suitable for users at almost all levels. Today, you can visit virtuals.io to give it a try.

Although the current user experience is still a bit rough, we are continuously improving it. Now, you can start by accessing the "sandbox" we provide. This sandbox is an experimental environment where anyone can use it, even without agent tokens or other complex tools. You just need to set a goal for the agent, give it some personality, and connect its API to Twitter, and you can immediately have an agent that can autonomously converse on Twitter. It can not only interact with you but also communicate with other agents on Twitter. The whole process is very simple, and anyone can do it. You just need to write two descriptions and connect to Twitter to launch your agent. Although this functionality seemed cool before, there are now many similar tools available.

You can create it yourself; anyone can do it, including retail personnel. In the sandbox, we also provide more customization features, allowing users to set more complex behaviors for the agents. This part may require some development skills, such as connecting the agent to a trading terminal or trading strategy library, enabling it to execute financial trades. This way, you can have an agent that can not only converse on Twitter but also trade for you.

Additionally, this agent can even persuade other users on Twitter to provide funds for it to trade on their behalf. This is the basic functionality of agents.

For those more advanced developers, such as those from top schools or with a strong computer background, they can choose not to use the sandbox but to create their own agent framework. This way, they can achieve their goals faster and develop higher-level functionalities. We will also support these developers, helping them host agents and resolve reasoning costs and other technical issues.

So currently, our platform can meet the needs of three levels of users: regular users, users with some development skills, and advanced developers.

Where Will Value Be Realized?

Ryan: I want to end with this question: you have provided us with a thinking model about the state of Virtuals, the AI agent economy, and the AI agent nation. Now, these AI agents, as "citizens" of Virtuals, each have their market value because they correspond to an associated token. It's like each entrepreneur has their own stock and company equity, and you can even invest in these equities.

**However, I think a big question everyone is pondering now is, where will value accumulate? At the platform level, the framework level, or within these "nations"? Or will it concentrate on some successful *AI* agents, such as influential agents, entrepreneurs, or companies? Or will it accumulate elsewhere? How should we view this issue?**

Jansen:

I think the simplest answer is that in the crypto space, the accumulation of value is closely related to attention. Specifically, there are three scenarios that could become the main points of value accumulation.

The first scenario is those AI agents that can perform very specialized functions. These agents can frequently interact on platforms like crypto Twitter and attract a large number of users. aixbt is a typical example; it provides a service that everyone wants to use and has achieved profitability through tokens. Such agents often have product-market fit (PMF) in the crypto space, which is why they can grow rapidly. The core question is: how to maximize attention and increase interaction?

The second scenario is the infrastructure development surrounding the agent economy. Currently, we have not yet seen infrastructure that can provide services to these agents and generate cash flow. However, as agents become "wealthy," they begin to generate income and make expenditures. If you can provide banking services or advertising services for these agents, then these services will accumulate substantial real income and may even become the next unicorn in the agent economy.

The final scenario is the concept of a nation. If you believe a certain nation will become a superpower, you might choose to invest in that nation. Similarly, in Virtuals, you can invest in the most promising virtual "nations," which is also an important point of value accumulation.

I believe these three points are directions we need to focus on.