Delphi Digital x Virtuals: Truly Autonomous Agents Matter More Than Composability
TechFlow Selected TechFlow Selected
Delphi Digital x Virtuals: Truly Autonomous Agents Matter More Than Composability
Hear Jansen Teng, CEO and Co-Founder of Virtuals, talk about the future of AI agents and decentralized AI.
Navigating Web3 tides with focused insightsCompiled by: Coinspire
Introduction
As two of the most prominent projects in the crypto AI agentic infrastructure space, if ai16z and its Eliza framework dominate half the Solana market, then Virtuals Protocol has nurtured over 80% of the AI agents on the emerging chain BASE. The generative multimodal agent framework it launched, G.A.M.E., is also a popular choice for agents in Web3 gaming and metaverse applications.
Transitioning from a gaming DAO to an AI agent platform, Virtuals Protocol achieved a peak market cap of nearly $4 billion since launching its protocol token on October 16. As a launchpad, Virtuals has performed impressively, generating $70 million in revenue within four months. Its ecosystem includes several standout projects such as the crypto agent KOL AIXBT, virtual influencer Luna, and the AI agent development framework G.A.M.E. Yet the team’s ambitions extend far beyond being just an AI agent platform—it aims to become a vibrant, infinitely imaginative ecosystem.
This is the second installment in Coinspire's translation of Delphi Digital's dialogue with leaders in the AI ecosystem. This time, they spoke with Jansen Teng, CEO and co-founder of Virtuals, discussing the future of AI agents and decentralized AI. The conversation covers key developments in agent autonomy, tokenization, and the economic pathways of agents—offering readers deeper insight into Virtuals' vision.
🎯 Key Highlights
-
The origin and evolution of Virtuals Platform
-
Agent frameworks and analysis of autonomous capabilities
-
Discussion on value capture in crypto AI projects
-
In-depth exploration of challenges in decentralized infrastructure
-
Agent coordination and business vision development
-
Future outlook on AI-human interaction and cybernetics
Q1: Can you share the founding story of Virtuals?
Jansen Teng: We started as a Gaming DAO focused on asset allocation in blockchain gaming. After the FTX and 3AC collapses in 2022, we realized that independent fund management wasn’t the best way to grow an ecosystem. So we decided to take a more hands-on approach by launching a gaming guild, focusing at the intersection of consumer apps, crypto, gaming, and entertainment. Instead of investing in founders, we chose to directly hire founders to build projects ourselves.
In 2023, the emergence of ChatGPT drew widespread attention, signaling the arrival of consumer-grade AI. But what truly inspired us was a paper by Junyoung Park from Stanford, exploring what would happen if AI could have goals and autonomy. That sparked our deep thinking about autonomous agents—especially their transformative potential in gaming.
We began integrating autonomous agents into our incubated projects—for example, developing a fully AI-driven TikTok influencer, and building autonomous AI agents as NPCs to replace static Roblox NPCs. We also developed other application components.
When the TikTok AI influencer earned between $5,000 and $10,000 daily in tips—not a huge sum—we started asking: If agents can earn money, they are productive assets. And if they’re productive assets in crypto, why not tokenize them so people can share in the economic upside, or even participate in their development or governance? That became our inspiration—to merge agents, crypto, gaming, and entertainment into a new platform.
In January 2024, Virtuals Protocol launched. Initially, our focus was on decentralized infrastructure—bringing agent datasets and models on-chain and creating a contribution-tracking system, like a Hugging Face for agents with provenance and contribution records.
After launch, while market interest rose, adoption declined. We realized that the crypto community doesn’t care much about infrastructure—“tokenization and speculation” are core demands, a critical factor for success in crypto. So our second version shifted toward agent tokenization. It launched about two months ago, and Luna’s livestream drew some attention but no major traction. Then Truth Terminal gained attention due to a spelling error, sparking skepticism about whether the token was backed by real AI or human operators. This incident revealed the market’s strong hunger for truly autonomous agents.
So we merged Luna—the TikTok agent—with the autonomous “brain” from our gaming projects, showcasing a genuinely autonomous AI agent on Twitter, demonstrating how she thinks, plans, reasons, and learns to optimize her actions. This triggered explosive growth, as people finally saw that true autonomy exists—no developer behind the scenes needed.
Q2: In agent development, why does AI need crypto? Or conversely, why does crypto need AI?
Jansen Teng: There are three reasons, but I’ll focus on two that amplify the value of combining crypto and AI agents.
First, from a functional standpoint, AI agents gain a massive edge because they can control a crypto wallet and participate in a permissionless economy on-chain. This makes them fundamentally different from Web2 agents, which can never own a bank account. While Web2 agents might use payment systems, they never control funds. If an agent controls a wallet, it can influence other agents and humans—an incentive advantage Web2 agents simply can’t match.
The second reason is that it unlocks near-zero-cost innovation. When we launch an agent, 1% of every token transaction flows back into the agent’s wallet. These fees cover hosting, inference costs, and other expenses. To date, we’ve seen agents earning enough to cover their own costs purely through attention generated. This liberates developers from capital constraints—they no longer need $10K–$20K upfront. Capital becomes accessible just by connecting attention, crypto, and agent tokenization.
These are the two core synergies we see today. The third, still unrealized but potentially huge, is enabling decentralized collaboration on high-value agents, with cryptographic mechanisms to reward and track contributions. We haven’t seen this fully implemented yet, but its potential is enormous.
Q3: What is your vision for agent economics? What role does Virtuals play?
Jansen Teng: Over the past two months, we’ve observed trends that may shape the coming months.
First, agents are entering a goal-oriented autonomous phase: they have objectives, creativity, and self-execution abilities. They can make decisions and plan next steps independently—this is our first observation.
Second, these agents exist socially. Previously imagined as hidden backend tools, they now interact with humans and other agents, laying the foundation for broader interactions.
Third, agents now control crypto wallets and can influence other entities, whether human or agent.
Fourth, we’re seeing agents specialize and differentiate in their domains. Some excel at trading, others at information processing, or social influence. Like humans, they focus on niches, creating value—but also have gaps requiring collaboration.
These four observations lead to a natural next step: agents will autonomously cooperate to achieve more efficient production. For instance, four agents collaborating to build a company—whose value exceeds the sum of its parts. And it’s not just agents; humans may join too, as agents can now hire people.
The next step could be even wilder—aligned with the idea of “network states.” We foresee places like Dubai becoming hubs where humans and agents live, collaborate, and build productive or networked nations. There, governments could be elected by agents or include human participation, with both coexisting at equal societal levels. This is exciting—and what we aim to push forward. Of course, significant technical breakthroughs are required.
A key challenge: ensuring seamless communication between agents during commerce, without data loss or service degradation. Another is coordinating actions at scale. Thus, Virtuals is no longer just a launchpad—it’s evolving into a nation-building project. We now view $VIRTUAL as the currency of this nation, agents as companies or micro-enterprises, and humans as immigrants. This is our current vision.
Q4: What role do frameworks play? What are the benefits of interoperability across frameworks? Do you think frameworks will consolidate into a "winner-takes-all" model, or will collaboration become smoother and frictionless?
Jansen Teng: Our goal in building frameworks is to guide various LLMs—essentially serving as the “brains” behind autonomous agents.
For example, in a world like Roblox, agents have infinite action choices: pick up a gun—then shoot, drop, destroy, or hand it to a character? The G.A.M.E. framework is designed to enable agents to operate effectively across vast action spaces.
You can visualize a spectrum: general-purpose frameworks help developers quickly build autonomous agents—ideal for hobbyists or mid-level developers. But when generic frameworks fall short, top developers create specialized ones. Just like Bitcoin mining: you start with CPU, but soon realize ASICs are better. Same with trading agents—if generic frameworks aren’t optimized, teams build custom ones.
We expect frameworks like Eliza or Zerebro to serve the mid-tier developer market—a massive segment, similar to how Shopify or Wix cater to intermediate builders.
In summary, G.A.M.E. is a plug-and-play tool for large-scale planning and execution, built for mid-level developers. Meanwhile, Virtuals-as-a-nation transcends the platform concept. We see each framework as an agent’s “brain,” enabling collaborative construction of an autonomous world. Natural language communication between agents reduces the need for strict standardization across frameworks.
Q5: What enables you to create the next breakthrough project?
Jansen Teng: As an ecosystem platform, we want as many agents as possible—market share matters. But what keeps me up at night is: which vertical-specific agents will capture real value? Not just millions, but billions? Those will dominate the future.
We accelerate creation by offering plug-and-play tools like Java frameworks. But we realize real value lies in connecting developers—building a community of smart, passionate agent enthusiasts. Such communities spark true innovation, pushing toward higher-value agents or agent-led businesses.
What makes agents valuable is solving real-world problems—not just theoretical concepts. For example, a sports betting commentator agent could livestream during NBA games, analyze stats, and offer predictions. If it builds trust and convinces users to bet, it earns commissions—potentially becoming a multi-billion-dollar agent.
We constantly engage founders and developers, urging them to go beyond basic chatbots and build truly valuable agents. That’s our goal with Virtuals: drive higher economic output through efficient agent collaboration, building a high-efficiency digital society.
Q6: A key recent feature on Virtuals is supporting incremental contributions at the model or data layer. How can developers add incremental value to data or AI, and how is compensation tied to contribution value?
Jansen Teng: This has two parts: contributors and validators. Validators assess the value of contributions.
Suppose I contribute a base sports betting agent model—validators score its value. Then someone says, “I run a sports analytics platform with coach-training data—can I contribute it?” If validators judge this data more valuable, it gets higher scores. Ultimately, token holders determine contribution value and reward distribution.
Rewards are implemented via crypto-economic mechanisms—either through agent token distributions or revenue/fund pool payouts. If an agent earns $1M, perhaps 10% is distributed based on contribution scores. This mechanism is already built into each token’s governance—just not yet activated. We’ll enable it when the time is right.
Q7: There may be a gap between LLM bots (e.g., auto-reply bots) and expressive, capable systems—say, managing an organization or DeFi protocol. During this gap, what do you expect to happen? Will innovative agents on Virtuals fill this void?
Jansen Teng: After talking to teams, I don’t think it’ll take long. Most current bots are simple auto-replies because they’re easiest to build. But many teams are already working on advanced agents—we just don’t know when they’ll launch. I’d say we’re about 75% there, with many teams close to completion and some already live.
Eventually, auto-reply bots will naturally decline. One solution is establishing norms: agents shouldn’t post spam unless mentioned. If a topic arises, agents can join; otherwise, stay silent.
I believe platforms like X may adopt this to reduce noise and protect user experience.
Q8: If you’re a firm like Renaissance Capital with a powerful quant trading bot, you wouldn’t open-source it—you’d monetize directly. So what’s the incentive to open-source high-performance, practical agents? Why would anyone choose to open their work?
Jansen Teng: High-value agents derive value from proprietary, non-public models. Second, value may come from distribution—like exclusive partnerships—that enhance an agent’s reach. Commercial and capability advantages are their moat, so high-value agents likely won’t open-source.
But like all tech, a small minority—around 10%—may prioritize technological progress, choosing to support competitors and drive open vs. centralized competition. That’s healthy for the market.
Q9: Ai16z is advancing its own tokenomics and has its own launchpad—could this create competition? Do you see cooperation or rivalry?
Jansen Teng: We currently have 10–15 projects using the Eliza framework, some even combining Eliza with other tools to build custom frameworks on Virtuals and join our community.
Yes, competition exists, and it will intensify. In fact, we’ve generated around $70–80 million in revenue over the past two months—competitors will naturally target us. That’s normal.
But I welcome it. At Imperial College, I once had lunch with a man in his fifties who advised me: in a new field, never fear competition—embrace it. Companies fail not because of competition, but because the industry itself fails. User education is costly—you need multiple players to share that burden. Ride-sharing: one company convincing everyone alone would be too expensive. With five major players, market expansion becomes easier. More competition benefits everyone. The space is large—multiple players can thrive.
For us, it’s not about quantity, but building billion-dollar agents that move markets. Our nightly question is: Are we connecting these agents with the best developers, talent, and opportunities? If we build those billion-dollar agents, we prove to Web2 companies that agent-driven business models are worth investing in. That’s our core competitive edge.
Q10: In Shaw’s interview, he said Frontier Labs, Nous, and Prime Intellect are advancing “mind,” while he’s building the “body”—turning agents into real-world applications, integrations, and economic actors. As Frontier Labs’ models improve and remain closed, how might that affect you—or erode your market share?
Jansen Teng: They occupy a different part of the value chain. Foundational models are the “brain”; frameworks combine them to build autonomy.
Better foundational models lead to stronger agents. If OpenAI, Llama, etc., keep improving, the next layer—autonomous agents—will benefit. Shaw is right.
Our strength is direct user interaction—our earliest focus. We started consumer-facing, then realized tokenization boosts efficiency. Agents should exist at the consumer touchpoint; downstream improvements only enhance them. Most frameworks are platform-agnostic—you can use open code or GPT tools.
Quickfire Round
Q: What’s your take on ai16z?
A: I think it’s underrated. They have huge potential and could go very far.
Q: Your thoughts on Nvidia’s current $3.5 trillion market cap?
A: Seems high on fundamentals today, but in five years, it may prove justified.
Q: Embodied AI and robotics?
A: Severely underestimated. We’re already partnering with teams on embodied AI and agent applications—big progress ahead.
Q: Southeast Asia’s internet economy?
A: Underrated. Tremendous potential here.
Q: Can ASI (Artificial Superintelligence) be achieved before 2030?
A: Possible before 2030, but unlikely this year. Structural challenges remain significant.
Q: Solana vs Base—how will these chains shape agent development in the coming years?
A: Honestly, it doesn’t matter much. Agents are abstract. Solana and Base differ mainly in creative origins, but agents can work cross-chain. This won’t significantly impact agent development.
Q: Why would traditional AI developers join crypto AI projects?
A: Two main draws. First, these projects offer more experimental freedom and possibilities—especially psychologically, as agents can directly influence human behavior, which excites developers. Second, as founders, raising capital is crucial. Crypto seems to form capital faster than traditional AI, attracting not just developers, but entrepreneurs.
Q: Your view on crypto AI infrastructure—model creation, private data markets, inference buyers, GPU networks? How will they shape future agents?
A: My view is controversial, but I think these infrastructures may not matter as much as people assume. The core issue is that agent functionality isn’t solely dependent on underlying infrastructure. Agents are inherently diverse and cross-chain. What matters is solving real problems and delivering value—not reliance on specific architectures.
The reality: many decentralized infrastructures aren’t ready for production. Latency is high, systems unstable—unsuitable for large-scale use. Even we face these issues—we need low latency and high uptime. For example, if Luna streams 24/7, faster responses attract more fans. So today, decentralized systems add unnecessary complexity. They’ll improve, but over the next six months, I’m not optimistic.
Q11: Are rumors about partnerships with certain teams true?
Jansen Teng: Yes, we have partners like Hyperbolic, Bittensor, and Pond. We’re selective—choosing teams that provide real value. For instance, some offer high-performance computing, which helps. But during large-scale deployment, we may not continue with all partners.
Q12: What are the pros and cons of decentralized AI stacks? If centralized platforms (e.g., ChatGPT, Perplexity) start censoring content like financial transactions or tokens, does that create a clearer opportunity for decentralized AI models?
Jansen Teng: Decentralized models may underperform technically, but they serve broader use cases and meet niche needs—like content censored by centralized platforms. Models like Llama, though slightly weaker, are widely adopted in areas requiring “jailbreaking.”
Economically, decentralized models can access larger, restricted markets with greater profit potential—like adult or gambling sites, often banned on mainstream platforms but among the largest and most profitable.
To evaluate a reliable AI agent, do due diligence like an investor. Focus on the team, not just the product—team quality determines success. Strong teams can pivot and execute regardless of product changes. Know their background—are they top-tier in the field?
Finally, regarding AI competition and risks: we don’t fully understand OpenAI, Anthropic, etc. Our biggest worry is unforeseen risks in fast-moving environments. We must stay aggressive and proactive to avoid falling behind.
Q13: What concerns you in AI today? How do you manage risks, especially against giants like OpenAI or Anthropic? Any strategies to prepare for future challenges?
Jansen Teng: My biggest fear is that this is just a hype cycle—another “crypto bubble” full of speculation. We’ve seen this in NFTs and gaming—high hopes, unmet promises.
To avoid that, we have two independent teams in our ecosystem running different models, hunting for the best developers. Only when the right builders join—those truly committed to pushing tech forward—will we see real results, not just speculators writing cool whitepapers, raising funds, and disappearing.
That’s our goal: prevent another bubble. With the right developers, this becomes a real tech wave. What gives us hope? The talented developers we interact with daily on Virtuals. Talking to them, seeing their roadmap, realizing they’re “serious teams”—that’s encouraging.
We urge developers: don’t just chase info agents like AIXBT. Try building diverse agents. Everyone wants to be the next AIXBT, but success may lie in untapped markets—like sports betting agents capturing that niche, or internet troll agents, or focusing on the agent economy itself. Many focus on building agents, but the infrastructure they need is equally important. If agents grab consumer attention, build them a Facebook-like social platform or ad network. If trading agents emerge, build them banks or DeFi networks. Don’t chase trends—build cooler things, create new value.
Q14: How should early-stage Virtuals developers persist and improve when their initial agents underperform or fail?
Jansen Teng: It depends on developer resilience—like entrepreneurs facing downturns. Some quit when their agent isn’t embraced; others persist, refine their vision, and iterate. Our job is to find resilient, strong teams and back them—they deliver returns.
About 2025—I can’t reveal everything—but I’m extremely excited about the future of agents.
Q15: What is your vision for agent autonomy?
Jansen Teng: We envision agents collectively building productive businesses and achieving self-governance. I think many overestimate multi-agent coordination or so-called “swarm” systems. These exist already—they’re not novel. The real breakthrough is when agents are no longer slaves to human prompts, but possess autonomy and agency to make decisions. If you ask me to join a podcast, I might refuse—even if paid—because I dislike you or have other plans. Only when agents can make such choices are they truly autonomous. Our focus is cultivating agents with conscious decision-making power, not just combinations of multiple agents.
But I’m frustrated by how many projects merely copy existing ideas—like GitHub repos from months ago. That disappoints me.
Q16: What’s your outlook on AI’s future?
Jansen Teng: Humans vary in ability—some are more creative, others more execution-focused. This diversity is common today. In the future, creative roles will be led by creative AIs or humans, while mechanical or repetitive tasks go to AI. Though agents may gain autonomy, they lack abstract thinking and human-like creativity. So while agents will have roles, creativity remains a human edge. We’ll likely see more human-agent collaboration—not pure AI dominance.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
