Z Potentials | Exclusive Interview with Wenfeng Wang: How Sheet0 Raised $5M with Just Two People to Build the Google for AI Agents
Z Potentials invited Wenfeng Wang, founder of Sheet0, to give a talk.
Over the past two years, AI Agents have gradually shifted from early technical exploration into real-world deployment. Whether it is the general-purpose Agents driven by leading companies such as OpenAI and Anthropic, or automation tools tailored to vertical scenarios, the market is quickly validating one trend—data has become the core fuel powering Agent capabilities. However, the difficulty and cost of acquiring, cleaning, and organizing high-quality data remain a critical bottleneck preventing Agents from reaching large-scale application. Traditional methods of data collection either rely on engineering teams to build programs or on inefficient manual operations. In today’s world of fragmented and real-time information, these approaches can no longer meet the dual demands of accuracy and speed.
Against this backdrop, Sheet0 has chosen to enter from a track long overlooked yet holding explosive potential—providing both humans and Agents with “Level-4 real-time data collection and delivery capability.” Unlike approaches that depend on calling a single tool, Sheet0’s foundation translates user data needs into executable code, building an interpretable and traceable dynamic workflow system. Combined with a feedback loop driven by a data environment, Sheet0 is able to self-heal when execution errors occur. In internal testing, Sheet0 successfully collected the complete set of information for 294 companies from Y Combinator’s official website in just two minutes, and could further expand to capture founders’ educational and professional backgrounds as requested by the user. Both cost and speed were an order of magnitude better than other Agent products.
Founder Wenfeng Wang, a serial entrepreneur, brings nearly a decade of experience in AI, foundational software, and large-scale distributed data processing. His expertise in data engineering, context engineering, and composable system architecture gives Sheet0 a solid technical foundation. This not only enables the product to run faster and more stably but also gives him a rare ability to anticipate market rhythms and positioning ahead of others.
Sheet0 is not just a data collection tool. It aims to serve as a highly efficient aggregation layer of real-time data for super-individuals, knowledge workers, and Agent networks—essentially building a Google.com dedicated to Agents. In the next phase of AI applications, whoever can solve the full-chain problem of data—from collection to structured delivery—will have the chance to become the critical hub of the Agent internet. With its technical system and unique market entry point, Sheet0 is quickly approaching that position. Let us now step into the story of Sheet0 and its founder, Wenfeng Wang—let’s enjoy!
Chat is becoming the new front-end of this era. What limits the capability of this “new front-end” is the model’s access to real-time data. Whoever can supply real-time data to models will become the “new back-end” of the Agent era.
Sheet0 positions itself as a Level-4 Data Agent. In our context, that goal is “real-time acquisition and structured delivery of data.” Users only need to state their need; we can then collect and organize the data into tables directly ready for analysis or visualization. This is our definition of a Level-4 Data Agent.
Compared to other products on the market, our greatest differentiation lies in how we can deliver results that are both fast and precise—“100% accurate, 0 hallucination.” The key to achieving this is breaking tasks down into interpretable, verifiable, and traceable workflow code at the core, making the data collection and processing fully transparent and the results fully reliable.
If we think of Agents as “people,” perhaps it is even more fitting to think of Sheet0.com as “Google.com for Agents” (laughs).
Isn’t human existence itself the best proof of AGI’s existence? After all, humans are the outcome of natural forces operating under complex physical laws. Intelligence itself is a result of physical rules at work. There is even a term called the “psychological barrier shattering effect.” Nature has already given us an example, proving that intelligence is indeed the result of physical laws—so why should we doubt that models will continue to progress?
AI development is moving fast. Everyone feels FOMO, and often people are being swept forward by the tide. I keep reminding myself: when something starts becoming widespread, you must jump out of the waves onto the shore and think—what is scarce?
01 From Programmer to Tech Entrepreneur
Z Potentials: First of all, congratulations on Sheet0 completing a five-million-dollar round of financing. I’m curious—was your team really just two people?
Wenfeng Wang: Thank you. When we signed the term sheet, it was indeed just me, two full-time colleagues, and one intern. Now we have a few more.
Z Potentials: Could you share your past work and entrepreneurial experiences? In particular, after graduating, how did your explorations in AI and foundational software shape your understanding of your current startup project?
Wenfeng Wang: My career can be divided into two stages. I graduated in 2017. For the first four years, I wrote code; in the following four years, I started ventures. In 2019, I joined Horizon Robotics, where I was responsible for the data platform under Horizon’s AI platform, mainly focused on data storage and cleaning. In 2021, investment in foundational software was very hot. Since my expertise was in message queues—a type of foundational software—I joined a friend in starting a company, becoming a co-founder and CTO. By summer 2023, I decided to build in the Agent space and threw myself into the wave of LLM entrepreneurship.
At Horizon, I worked alongside brilliant algorithm engineers and scientists, and under their influence I came to realize that only high-quality data produces better intelligence. My long-term experience in developing and designing foundational software cultivated my technical taste and product intuition, which have had an extremely important impact on how I think about building products.
Z Potentials: You’ve spent your first six years dealing with data. How do you view the importance of data in today’s Agent-human interactions? And can you talk about the special value of data in the Agent era?
Wenfeng Wang: First, we must distinguish between types of data and their application scenarios—data for humans and data for Agents are completely different. For Agents, the value of data is reflected on three levels.
The first is model training. High-quality data is the foundation of model intelligence. The industry’s discussions around GPT-5 not meeting expectations essentially stem from the fact that high-quality training data is nearing exhaustion. In autonomous driving at Horizon, we also found that rather than fine-tuning parameters, adding a batch of high-quality data had a more significant effect on specific cases. Premium data must clearly define inputs, outputs, and intermediate process data; these are the true core of model learning. Reinforcement learning equally relies on this kind of task-step data and reward signals.
The second is during runtime. The core data here is context, which should include all the step data involved in task execution, managed structurally, processed semantically, and defined in terms of causality. Claude Code’s recent strong performance, for example, benefited from finely distinguishing the semantics of tool-use messages. For Agents, good data is well-organized, semantically rich context.
The third is tool use. Essentially, this is about accessing third-party real-time data to supplement their own context. Today’s model tool use, or MCP, is basically wrapping a code function and then using that code to access the underlying data. But if we think carefully—how necessary is that code function? The essence of code is manipulating structured data at its foundation. So why shouldn’t a model directly query data with SQL, instead of being forced through a code middle layer?
For human users, the value of data—besides its business value—lies more in verifiability and psychological security. Users need to confirm the source of results. Take web data as an example: how was the data extracted from the webpage, and through which processing steps did it pass afterward? Can we present this to the user in a simple, direct, end-to-end “white-box” way? Accuracy and interpretability, I believe, will be the scarcest qualities of data. We must ensure that users have the confidence to use it.
02 Sheet0, the First Level 4 Data Agent
Z Potentials: Please introduce to us, in essence, what the Sheet0 platform is and what functions it provides.
Wenfeng Wang: Sure. I’d like to explain it from both a short-term goal and a long-term vision. We position Sheet0 as a Level-4 Data Agent. The term “L4” is most often used in autonomous driving—L4 means that you only need to tell the car to go from point A to point B, and it will handle the entire process itself. Applying the same logic to Agents, the user only needs to describe the goal, and the system will autonomously complete the task and deliver the result.
In our case, that goal is “real-time acquisition and structured delivery of data.” Users only need to state their requirement, and we can collect and organize the data into a table directly usable for analysis or visualization. That’s our definition of a Level-4 Data Agent.
In the short term, our core capability is converting any data source—webpages, files, APIs—into a structured Data Sheet. Let me give you a real example: one of our clients wanted to analyze which KOLs on Twitter had a higher “viral rate.” The logic is simple—Twitter is an important channel for public opinion and marketing, so when investing in campaigns, naturally you want to choose creators with higher viral rates to maximize ROI. The client simply told Sheet0 this goal. Sheet0 automatically collected the historical tweets of relevant KOLs, along with their view counts, replies, retweets, and other engagement data. It then structured this into a table and calculated each KOL’s viral rate using SQL, ultimately delivering a complete result.
Compared with similar products on the market, our biggest differentiator is how we deliver results “100% accurate, 0 hallucination,” while being both fast and precise. The key lies in breaking tasks down into interpretable, verifiable, and traceable workflow code, making the data collection and processing transparent, and ensuring reliable outcomes.
In the long term, Sheet0 will provide super-individuals, knowledge workers, and Agents with linearly scalable real-time data acquisition and processing, all through natural language.
Z Potentials: How should we understand this? In your view, why is data so important to Agents in the future?
Wenfeng Wang: In software, the back end essentially exists to provide the front end with real-time, renderable data. That data is scattered across different databases, file systems, or APIs, so back-end engineers need to manually aggregate it. I often think about—or even question—things that many consider common sense. For example: do users really need a universally defined front end? My view is that Chat itself is the new front end, and single-use apps are generative UI components that appear within a dialogue box. From GPT-5 onwards, I believe we will very quickly see this trend.
What limits the capabilities of this “new front end” is the model’s ability to access real-time data—because the “new back end” has not yet appeared. Whoever can supply models with real-time data will become the “new back end” of the Agent era. From this perspective, Sheet0 has the opportunity to become that “new back end.”
Why is real-time data so crucial? For Agents, data can be divided into knowledge and information. Models provide Agents with sufficient knowledge, but what they severely lack is information to make real-time decisions. The essence of Context Engineering is organizing this information for Agents—partly memory, and partly real-time data. But the ways Agents currently acquire real-time data are still very primitive and unintelligent. We need a more standardized way for models to access real-time data.
This also means that the competition in the future may not be about “how much data I own,” but about “how fast and how precisely I can provide real-time data to Agents.” Once data acquisition approaches real-time and transaction costs approach zero, the efficiency of collaboration between Agents will undergo a qualitative leap. Imagine an Agent performing a task, instantly calling multiple sets of data from different nodes—just as naturally as today’s API calls. This model will drastically compress information asymmetry and the capability boundaries of Agents.
So, our focus is not on becoming the sole “data source,” but rather on becoming the efficient data aggregation layer within the Agent ecosystem, providing real-time data access in a standardized way. If we think of Agents as “people,” then perhaps it’s more fitting to think of Sheet0.com as “Google.com for Agents” (laughs).
Z Potentials: How do you ensure data is truly “100% accurate, 0 hallucination”?
Wenfeng Wang: Many Agent products today emphasize autonomy, aiming to complete tasks end-to-end for the user. We, however, adopt a “multi-step confirmation and progressive alignment” strategy, with accuracy as the priority. So our “100% accurate, 0 hallucination” applies only to data that Sheet0 has successfully collected and delivered to the user. If the task fails, the user receives an empty sheet. There’s no uncertainty in between.
In terms of implementation, we mainly rely on two things.
First, dynamically generated and iteratively optimized execution workflows. We don’t hardcode a fixed scraping logic. Instead, when the Agent encounters exceptions during execution, it rewrites and optimizes the workflow and code. This means that the underlying logic may be completely different across multiple runs of the same task, but the user is unaware—what they see is simply a completed data table.
Second, a Data Environment–driven feedback mechanism. The Data Environment is Sheet0’s RL environment, supporting runtime monitoring, error classification, context filtering, and data validation. This ensures that after every failure, the Agent adjusts based on clear, structured feedback, rather than guessing through vague language descriptions.
From a technical standpoint, Sheet0 is essentially a Coding Agent highly skilled at solving data problems—writing code, running tests, collecting errors, then iterating improvements until the workflow can operate stably over the long term.
Z Potentials: You already have some internal beta users. Can you share a few user scenarios? Have you categorized your users, and are there any surprising “long-tail” cases where people found Sheet0 particularly effective—even solving major problems for them?
Wenfeng Wang: Our main use cases right now are focused on collecting public internet data.
The first case comes from an overseas outsourcing platform—sites like Fiverr, where “data collection” has always been a big service category. One U.S. user posted a task with a $20 budget: scrape event information from a Taiwanese online event-sharing website. The site was extremely outdated, built in a style from 20 years ago. The event data wasn’t presented in a table, but scattered as markers on a map. The user’s goal was to filter through over 700 events, find those suitable for children aged 2–4, and compile them into a list.
Traditionally, this would require manually clicking each map marker, opening a popup, and copy-pasting details—highly time-consuming. Because the data was tied to map interactions, hiring a programmer to write a script would be costly, slow, and unlikely for a low-budget task. Our tool perfectly addressed this pain point.
The second case is a classic sales lead generation need. A user was searching for potential AI customers, and their workflow had two paths.
The first was collecting information on AI companies from public sources. They would scrape company lists from AI tool directories, aggregate and organize them, and then find the right contacts and emails for each company, scattered across official websites and public channels.
The second was analyzing existing large sets of customer emails to see which companies were undergoing AI transformation. They would map emails to companies, study their industries, and tag whether they were related to AI.
Traditionally, building such an end-to-end pipeline with software development would take programmers weeks. But this user, with Sheet0, directly linked data collection, cleaning, matching, and labeling into a complete workflow, quickly producing a usable lead list.
During our waiting list phase, we did almost no promotion—relying solely on organic growth—and accumulated 3,000 waitlist users. We plan to open early access testing on August 12.
Since the ultimate goal is still far ahead, the path we take is critical. By aligning product direction with real user needs, our initial focus is on public internet data scraping—first becoming the strongest AI-powered data collection tool, one that can turn any public website into a structured data table. We are beginning by deeply supporting the needs of product managers, marketing operators, and others who rely heavily on data.
Z Potentials: And in terms of your business model, how are you planning? Especially regarding AI taking on tasks for humans, what kind of community do you envision building in the future?
Wenfeng Wang: In the short term, we’re using a credit-based billing model, similar to most Agent products, calculated based on underlying resource consumption. In the longer term, we’ll explore pay-for-result models. For example, if the data a user needs cannot be found in public channels or our existing databases, the user can set a price they’re willing to pay, create a task, and distribute it through the Sheet0 user network. If another user happens to have that data and is willing to sell it, a direct transaction can occur.
The key here is that pricing power rests with the user. Sheet0 knows who has the data and acts as the matchmaker, enabling data to flow efficiently and controllably between users.
03 From Experimentation to Long-Term Vision
Z Potentials: After you started your company in 2023, you also went through some directional pivots. In terms of industry judgment or technical route selection, did you make any detours? And what did you gain from those experiences?
Wenfeng Wang: Overall, I’d say I was too impatient and didn’t manage expectations well. I officially started my entrepreneurial journey in July 2023. My first direction was AI Coding, and the company was directly named LLM Programming, meaning “large language model programming.” Our entry point was even more aggressive than Cursor’s at the time—we wanted users to first write out a product requirement document (PRD), and then directly turn that into usable software. In essence, this was the same as what we now call Vibe Coding.
As many people may know, before May 2024, Cursor’s performance was fairly mediocre until Claude 3.5 came out, and then it suddenly took off. Back then, our understanding of and trust in models was not yet deep enough, and we pushed forward too quickly. The results fell short, so we decided to change direction. Looking back, that experience taught me to start believing that models will indeed keep improving.
The second direction we pursued was Tool Use. We called it NPi, short for Natural Programming Interface. It even made it into the top ten on Hacker News and was featured in the world’s first Agent Infra report published by Madrona Ventures. We launched NPi in February 2024, and by June we had a relatively complete first version.
The reason we chose Tool Use was that the first attempt convinced us Agents must be able to take action. In fact, looking back now, what we built with NPi was essentially the later MCP—we were six months early. We not only defined the processes and specifications but also built a complete implementation. Around the same time, there was another company, Composio, an Indian team. They completed a $29 million Series A led by Lightspeed US earlier this year, and now they’re one of Silicon Valley’s rising stars.
After we finished the first version of NPi, I spoke with over 30 Agent developers. I found that none of them were focused on Tool Use. (In fact, MCP only gained real traction in February 2025, after Cursor adopted MCP.) Instead, everyone was working on RAG. The core reason was that the problem wasn’t the absence of suitable tools—it was about getting the Agent to pick the right tool. In that process, I realized the key to Agents lies in context, which led me to formulate the concept of Context Engineering.
If you only build on the Tool layer, you don’t have access to the Agent’s context, and thus you can’t improve tool-calling accuracy. That’s when I realized that simply making a Tool Use product wasn’t enough. So we decided to move up a layer. After a long period of exploration, that’s how we ultimately arrived at Sheet0.
From the first to the third project, the underlying intention hasn’t changed: enabling ordinary people to gain engineers’ “superpowers”—the ability to simplify complex problems and then automate them. This has always been my guiding principle throughout each pivot. At the start, I thought: let AI help people write code. Later, I realized code is still unfriendly to non-programmers, so we tried Tool Use, because what programmers essentially do is link different code tool interfaces. If models could handle Tool Use well, perhaps people wouldn’t even need to see code. Finally, I realized the essence of code is simply operating on structured data. So there’s been a natural through-line.
In summary, my judgment usually leads the market by about a year. This means that to stand a chance, I need to endure at least half a year of negative feedback while grinding away. For example, both AI Coding and MCP required about a year before the timing was right.
Z Potentials: Yes, even if the direction is correct and the demand is real, it still takes at least half a year of grinding before you start to see some positive signals. AI isn’t a one-wave business—it requires long-term investment and accumulated momentum, with multiple bursts along the way.
Wenfeng Wang: Exactly. If we look back, there are almost no breakout products that emerged within six months of being founded. Those star products we see today—every single one of them had to sit on the “cold bench” for a year or two. Take Cursor, for example—it was founded in 2022, but only began to really take off after May 2024 when Claude 3.5 was released. Or look at Manus. Before that, Monica had already spent two years accumulating experience, and even after deciding to build Manus, the journey wasn’t smooth. There were over six months of adjustments along the way.
So my summary is this: first, you need reasonable expectations—especially not underestimating the engineering complexity of building production-ready Agents. Just a couple of days ago, I saw a diagram that I felt captured this perfectly. Second, you have to maintain faith in the continuous progress of models. It’s a cliché, but every time a new model comes out, there are always voices of pessimism. These days, I simply block them out.
My turning point of confidence came in May 2025. Before that, I had been forcing myself to believe. Then, one day, while chatting with a friend, a thought suddenly popped into my head: Isn’t the existence of humans themselves the best proof that AGI can exist? Humans are the product of natural forces operating under complex physical laws. Humans have intelligence. From this fact, it follows that intelligence itself is a result of physical laws at work.
There’s a term called the “psychological barrier shattering effect”—once a problem or limit has been proven possible, the difficulty of reproducing it drops drastically. The marathon is a classic example. Returning to my earlier logic, which may sound a bit odd: nature has already set the example, proving that intelligence is the result of physical laws. So why should we have any reason to doubt that models will continue to progress?
Z Potentials: Finally, what would you like to say to readers who are now starting their own ventures in AI?
Wenfeng Wang: AI is developing fast, and everyone feels FOMO. Many times, people are just being swept along by the wave. I keep reminding myself: when something starts becoming overhyped, you must jump out of the wave onto the shore and ask—what is truly scarce? I hope I can spend more time reflecting on that in the second half of this year.
Disclaimer
Please note that the content of this interview has been edited and approved by Wenfeng Wang. It reflects the personal views of the interviewees. We encourage readers to engage by leaving comments and sharing their thoughts on this interview. For more information about Sheet0, please explore the official website: https://www.sheet0.com/.
Z Potentials will continue to feature interviews with entrepreneurs in fields such as artificial intelligence, robotics, and globalization. We warmly invite those of you who are hopeful for the future to join our community to share, learn, and grow with us.
Image source: Wenfeng Wang and Sheet0