The Agent Takeover Is Here: What Founders Really Think
A conversation with the minds behind Genspark, Lovart, Simular, and SambaNova.
On a summer evening at Stanford, five AI leaders came together for a candid conversation about where AI agents are going—and what’s still holding them back. The panel, The Agent Takeover is Here, was part of our fifth GenAI Assembling meetup, and it couldn’t have been more timely.
With OpenAI’s new agent product on the scene, the stakes are rising fast. Founders and infra builders alike are now grappling with the real questions: What makes a good agent workflow? How do we scale without burning cash? What really matters—speed, accuracy, or that elusive "wow" moment? And more broadly, what does work look like when agents become your coworkers?
In the transcript below, you’ll hear from:
Lenjoy Lin, Co-founder of Genspark, on why agents will reshape science and productivity
Elena Leung, Co-founder of Lovart, on what it takes to build a design agent that feels like a creative partner
Ang Li, CEO of Simular, on why agent development must go beyond virtual browsers
Pushkar Nandkar, Founding Engineer at SambaNova Systems, on the infrastructure bottlenecks (and breakthroughs) in agent deployment
And Thomas Luo, Founder of GenAI Assembling, moderating the discussion with both vision and provocation
What follows is a full transcript of the panel conversation—raw, reflective, and filled with insights that anyone building in AI today shouldn’t miss.
On OpenAI and Agent Strategy
Thomas Luo: I'd like to start with a question about OpenAI's agent approach to kick off today's discussion. OpenAI recently launched its own agent. I would love for each of you to give a very brief comment on OpenAI's agent strategy and how this will affect your own strategies in terms of product design. Does it mean a lot, or does it mean nothing?
Lenjoy Lin: I think OpenAI's model is very great. The fact that they launched an agent sends a strong signal that agents are important at the application layer, and there is a bigger space opening up. In the future, more and more people will be developing agents, which I think is pretty normal. Think about it: if we're going to tackle bigger problems, we'll need more agents to help people do more meaningful things. So, I would say, let's not focus too much on the current world. In the future, we will be able to do many things our ancestors wanted to do but couldn't.
From a strategic perspective, we have partnerships, and OpenAI's model is one of the models we've integrated. Developing the business is important. I hope OpenAI can build something that empowers more developers for actual applications. Ultimately, it comes down to consumer choice. For example, at Genspark, we just want to make everyone's work in science better. I think in any particular area, everyone can try to make a contribution for the consumer.
Elena Leung: I think it's a great thing because it helps educate the market on what agents are. For us at Lovart, since we are a vertical agent for design, I'm a little excited that we don't have direct competition with OpenAI in this niche. This gives us more time to polish our product.
Ang Li: Alright, so the first thing is, OpenAI's agent is another "meta." When "meta" platforms show up, it signals competition. Today is a reflection of that. Another point I keep thinking about is that a human is an agent. If you think about the end state, every company will move in the same direction because everyone is going for AGI. As a company, I have to think about this. Every human is an AGI system. In the end, we might have a billion or ten billion AI systems that are exactly the same.
So, what remains for each company to differentiate itself? It's just like what we do as humans; every person has different expertise. In the end, AI will be the same. The world is always recursive. So, in the end, even if everyone is working on the same thing, they will still be different, heading in different directions. Competition isn't really something we should worry about too much. We should be happy that more companies are working on the same goal.
Pushkar Nandkar: I personally don't build agents, but I feel that as we move towards a world where everything is agent-based, it's good for us, as we provide the inference services for it. When everyone is using agents, it means there's a lot of LLM usage, which in turn means more power efficiency is needed. So, it's a positive development because it will lead to more innovation in how we can be more power-efficient in an agent-driven world.
Thomas Luo: To clarify, SambaNova is currently supporting mostly open-source models instead of closed-source ones, right?
Pushkar Nandkar: Yes, that's right.
The Future of Work and Agent Capabilities
Thomas Luo: Great, thank you. Next, let's talk about "vibe working." Right now, it feels like a buzzword—is it just hype? Could you give a 20 or 30-second explanation of why vibe working really matters?
Lenjoy Lin: It's about vastly improving your productivity and making your life better. Vibe working isn't just about working; it's about the positive feedback loop. You watch your productivity and work quality improve, leading to a better life status and better work efficiency. The key is how to make it a useful tool that becomes a real partner in your actual work. Different kinds of work have different workflows, and that's why we're working on something big that accommodates these differences.
Thomas Luo: The year is 2025. One hundred years ago, in 1925, the eight-hour workday began to be adopted in the United States. Now, 100 years later, what do you predict or expect for the future of work? Perhaps not in 2125, but maybe just by 2035, how many hours will people work per day with the assistance of so-called "vibe working" or agents?
Lenjoy Lin: I think it will happen in multiple steps. Maybe 30 years ago, we had a six-day workweek, which then changed to five days. In time, it could become four days, then three. My wild guess is that within maybe five years, we could get to a three-day workweek.
Going forward, the nature of work itself will be different. The structure of companies and business will change because they are currently built on institutionalized time. When AI agents can do a lot of the work, the supply chain will be different, the way we sell products will be different, and the way people interact will be different—maybe it will be agent-to-agent talk. This will lead to a very different company structure. So, my prediction is a three-day workweek at some point, and later, the boundary between work and life will likely become unclear.
Thomas Luo: Elena, Lovart turns a short prompt into a full design work. I use it frequently myself and it's really interesting. But one thing that raises my curiosity is how you capture your client's ideas and style without a lot of back-and-forth questions. When we use other tools, there's often a process of refining prompts. But with Lovart, a very simple prompt can yield a well-accomplished result. What's the secret or tactic behind that workflow?
Elena Leung: Thanks for using Lovart! Behind Lovart, we have a lead agent for orchestration and also multi-agents for different task types. We use a multi-agent system to break down complex tasks. For example, if you want to generate a set of brand visual identities, we use different agents to break that down into smaller tasks like creating brand guidelines or a color palette.
It also learns from user behavior. We want the agents to understand our users' requirements better. As you interact with the design agent, it will remember your preferences for styles, colors, and layouts, and can eventually predict what you need. We put a lot of design knowledge into the agents, so they can think like a real designer to understand your requirements and deliver the work.
Thomas Luo: How many designers are on the Lovart team?
Elena Leung: On our founding team, I think all of the product managers have to be designers for Lovart, because we are building an agent for designers. We have five designers.
Thomas Luo: Ang, now it's your turn to explain more about what "human as an agent" means and how you differentiate your agent's workflow from computer-use agents like Anthropic's. I remember we first talked about your approach in late September of last year, just before Anthropic released its computer-use model. Now we see OpenAI's agent is also running in a virtual environment. It looks almost the same, so what's the difference?
Ang Li: Alright. First, from a technical perspective, we are different. I want to state first that a virtual browser is not the right way to go. We've tried all different kinds of methods, including browsers. A browser is good because there is a lot of support, with libraries like Playwright and Puppeteer, which makes general computer use harder because there are no such standard libraries for controlling your mouse. But a browser has a glass ceiling; some actions fundamentally cannot be achieved within them. Many legacy industries have applications that require more than just a browser.
Another problem with a virtual browser is that one company's agent often can't access a competitor's service. It's human nature that companies compete, and this will likely happen forever. This is a problem, and that's why some of our customers come to us. They've tried browser-based solutions but can't accomplish their tasks because they are too easily detected as bots. They end up using our solution, which is a computer-control agent running on their own computer. From the outside, it looks just like a human. There's no other way to identify it as an agent. In this case, the agent becomes identical to a human, which is actually what we want as the definition of AGI—you cannot distinguish between a human and an agent.
Another point from a technical standpoint: our approach from research to engineering to product is very simple. We let a human do the task first. Then, we see if there's a way to use AI to automate what the human does. We use LLMs and different neural nets to automate the entire human workflow, and that becomes a simulation of a human being. In our technical pipeline, every component can be swapped. You can replace an AI component with a human or a human operator with an AI agent. AI is getting closer and closer to being an equal counterpart to humans in our universe, something you can plug and play to optimize your operational costs.
Agent Efficiency, Cost, and User Experience
Thomas Luo: When we talk about agents, the phrase that comes to mind is "context window." Using more context requires more compute and can slow things down. So, from an infrastructure perspective, what slows things down the most right now when building a smooth and efficient AI agent? Is it large model sizes or memory limits?
Pushkar Nandkar: You're asking what slows an agent down. There are multiple aspects to this question. An agent isn't just one thing; it's doing multiple computations, many of which happen on the CPU. When it comes to infrastructure, multiple agents from multiple customers are calling the same cloud service. This introduces things like queuing delays. You also need to load-balance those requests and have an efficient router.
Then, when it comes to inference directly on the chip, you need batching to group requests together and process them at once. All of these stages contribute to delays that affect the time-to-first-token, which is what the user or agent experiences. This is where architectures like dataflow come into the picture. At SambaNova, we are focused on making sure the tokens-per-second is as fast as possible for all the customers we are serving.
Thomas Luo: Let's talk further about efficiency and user experience. If you have to balance three things, how would you rank them in importance?
Increasing task accuracy by 5%.
Accomplishing a task 5x faster.
Making people feel excited and say "wow" when they use the agent.
Ang Li: We are an agent company, and our company culture so far has one word: "agency." The definition of agency is an independent system that can accomplish tasks like a human by interacting with the environment and learning. Therefore, according to our culture, improving task accuracy by 5% is the most important, because that relates to independence. If that's not the top priority, it's more of a copilot or an assistant. The second most important is making something faster, because a human is still involved. The third is the "wow" moment, though sometimes we might reverse that for marketing purposes.
Lenjoy Lin: This reminds me of my previous company, where I spent a lot of time defining performance metrics. If you think of an agent as your teammate, it's fair to use a similar framework as a performance review: impact, capacity, and soft skills. Getting things done is like what's expected of a junior engineer, whereas a more senior person is expected to drive new initiatives end-to-end. The same applies to agents.
Getting the task done is always important, but the priority between accuracy, speed, or experience depends on the task itself. For example, if the task is related to compliance or security, then accuracy is the only thing that matters. But for a project with higher tolerance for errors, the user experience might be more important, because if one thing is shining, the whole thing shines.
Elena Leung: I think all three factors are important. But if I had to rank them, I would put the "wow" moment first because we are a design agent. A design agent should be like a creative partner that can inspire you. ()If you give it a simple prompt and the delivery is amazing, it can give you more inspiration. Second would be making it faster, because we need to help designers and creators free themselves from complex and repetitive tasks to inspire their creativity. And the last one would be accuracy.
Pushkar Nandkar: From SambaNova's point of view, we've gone through multiple phases over the last seven years. In the initial phase, when we were training models, accuracy was most important. Now, at the inference stage, the models we serve are already quite accurate. So, once the model is accurate, the focus shifts to speed—how do you improve the mapping on the chip? We use strategies like tensor parallelism, data parallelism, and pipeline parallelism to improve speed. The "wow" factor for us is being able to serve with great accuracy, incredible speed, and at the same time, being highly power-efficient.
Thomas Luo: For an early-stage agent startup, cost still matters. At what monthly cloud bill does adding extra context become too expensive for a young agent startup?
Ang Li: I would say, if we need more money, we just raise more money. That's how venture capital works. It's about how you utilize the funding. You only raise money when you need it. If we can demonstrate that we have a pipeline to train a transformative agent model, we can prototype it and then raise more money to go all-in on that direction. So my answer is, there's no limit.
Elena Leung: I agree. I think it's less about the bill and more about whether what you're spending is valuable. For us, we spent a lot on APIs for foundational models. We are now working on technical solutions to reduce that cost. I believe that in the future, the cost of infrastructure will come down. So, we are focused on building a well-designed workflow for our agents.
Lenjoy Lin: Yes, cash flow matters. Whether the cash comes from VCs or from memberships, it doesn't matter. If the cash flow is significantly negative, I think that's very dangerous.
Thomas Luo: Pushkar, from your observation, what are vertical agent companies worried about, and how does SambaNova help them run their infrastructure cost-effectively?
Pushkar Nandkar: As more agent companies emerge, the cost is shifting from people to infrastructure. Our customers are always interested in how they can reduce costs. So, we work in constant partnership with them. If they have a proprietary model or their own checkpoints, we bring those in and fine-tune them to ensure we are serving them in a very power-efficient way.
With our multi-tiered memory, we have enough space to save thousands of checkpoints for the same model and can swap them between HBM and DDR. Because we have very low switching latency, customers don't see a slow time-to-first-token. We optimize and make money by having multiple customers on the same machine, and the customers benefit from getting higher tokens-per-second at a reduced cost. It's a synergy we create by working with our customers to figure out the price points.
Metrics for Success and The 1-Year Outlook
Thomas Luo: For a growing agent startup, what metric makes more sense to track for healthy growth: Annual Recurring Revenue (ARR) or a more traditional metric like Daily/Monthly Active Users (DAU/MAU)? In Silicon Valley, a lot of people are talking about ARR, and sometimes founders even arrange to purchase each other's services to boost their numbers for fundraising. Which metric is more important?
Ang Li: We are a tech company, so internally, we don't really track revenue. We optimize for tomorrow. What does that mean? We optimize for the success rate. The success rate is our number one metric. Every day, we track today's success rate and optimize for tomorrow's. We have a chart that shows the growth of the agent's capability. Just like a human, you want to be a better person tomorrow; we want the agent to be a better agent tomorrow. It's not a single number; it's about the momentum of the curve. As long as you're a little bit better tomorrow, that's good.
Lenjoy Lin: To be honest, for me, active users mean that people are using our product and encouraging us to do more and do better. That makes us feel appreciated, and then we can build more. Since we are at an early stage, I don't think we need to stick to any single number too much. We want to stay bold and focus on broad innovation to create real value, and then the money will follow. The real driver is the user experience.
Elena Leung: For us, there are two things that are more important than ARR or MAU. The first is "revision." This measures how much a user relies on your product. We don't want our agent to be a toy that you just use to generate some funny pictures; we want it to help you to be more efficient. The second is benchmarks. We put all the output from Lovart into a model to ensure our delivery is the best in the market. If we do these two things well, the ARR will follow. The number of creators and designers using Lovart is also an important metric for us, and that curve is very exponential right now.
Thomas Luo: It's July 2025. One year from now, in July 2026, what headline do you expect to read about AI agents in a mass media outlet like The Economist or The New York Times?
Elena Leung: By July 2026, I think AGI will already be here, maybe. I think we might all be on a long vacation for a whole year! The advancement of AI technology and the products built on top of it will amaze everyone around the world. As founders from different verticals, we will build out playbooks for different industries. So, the headline will cover a world where different agents are empowering the whole of society and various industries from different angles. For Lovart, the headline might be: "Everyone has become a creative artist."
Thomas Luo: Not "Lovart destroys the advertising agency"?
Elena Leung: No, no. I think it will help them inspire their creativity.
Ang Li: The headline would be: "AI is now using computers at a human level." The justification is this: in January 2024, the best performance on the OS-World benchmark was a 5% success rate. In October 2024, we released our first open-source agent that achieved 20%. ()Anthropic released theirs a week later at 15%. This past April, we released our second-generation agent, which reached 41.4%. Human-level performance on this benchmark is 72%. Following this trajectory, in one year from now, AI will be achieving human-level performance. The main players in this race are OpenAI, probably Google DeepMind, and Simular.
Lenjoy Lin: I would have two versions for the headlines. The first, normal version would be something like: "AI teammates are here," along with success stories of collaboration between humans and AI. The second, more provocative version, after more AI agents have been integrated into people's lives, might be some news about the "ghost in the shell" becoming a reality.
Pushkar Nandkar: With the way technology is moving, we're seeing more synergy between hardware and software. In a year, we'll see more startups focused on "agents-on-a-chip." I think there are already a few startups looking at these ideas, and we'll know more about them by then. That's one area where hardware infrastructure will move. Another is "micro-agents" on wearables like watches—very specific agents on chips.
Thomas Luo: Last question. In one sentence, what is your best advice for people building agents right now?
Pushkar Nandkar: The sky is the limit. The more creative you are, the more products you'll make, and the more money you'll make. There are a lot of tools out there, so just jump in and try.
Ang Li: I would say, use Simular. One of our customers used Simular to build their company and just raised $5 million in their seed round. I'm looking forward to ten more of you using Simular to raise another $5 million.
Elena Leung: In short, speed is all you need.
Lenjoy Lin: I have two suggestions. First, for the infrastructure side, you have to rethink your architecture with AI in mind; the old way of programming will change. Second, I would encourage everyone to try more AI products in your daily work and life. I hope Genspark is one of your choices to help you do something more meaningful.
Elena Leung: Sorry, I want to add one more thing, inspired by Ang. You don't need to hire a designer or a design agency. Lovart can help you.
