ViaBTC CEO Yang Haipo: From Nof1 to x402, a brief discussion on the application and future of AI Agents

With the Nof1 AI live trading competition and Coinbase's x402 protocol gaining widespread attention, the application scenarios of AI agents in the financial and payment fields are constantly expanding. As a representative of AI payment protocols, what are the differences between the x402 protocol and traditional payment protocols? What are its payment scenarios? And as AI payments mature, what foreseeable application scenarios will AI agents have in the future? To address these questions, we invited Yang Haipo, founder and CEO of ViaBTC, to have an in-depth discussion on the feasibility of the x402 protocol and the future potential of AI collaborative networks.

Q: The x402 protocol has recently become a hot topic in the industry. What are your thoughts on the x402 protocol's use of token payments to solve payment problems in AI?

Yang Haipo: Actually, putting aside the concept of "AI payment", from an engineer's perspective, x402 is a relatively simple protocol. Its core is not to invent a new payment method, but to package on-chain payment into a standard web service and solve the trust and execution problems of on-chain payment by introducing a Facilitator.

Many people compare x402 to traditional payment methods, but they actually serve different target audiences. Traditional payment systems like Alipay and Visa offer excellent user experiences, but they are designed for humans, not AI. For AI agents, traditional payment systems currently present two significant problems: First, the barrier to entry. It's difficult to open a bank account and complete KYC using a script, but generating a wallet address capable of on-chain payments only requires a single line of code. Second, there are transaction costs. AI interactions are frequent and fragmented. For example, if an agent calls a data interface and incurs a $0.0001 fee, the bank's transaction fees might exceed that amount if using Visa's network.

Therefore, x402 actually leverages the programmability of tokens, combined with the middleware role of Facilitator, to solve the problem of this "automated micro-payment" scenario. In this scenario, Facilitator is like the "Alipay" of the machine world, handling the complex on-chain confirmations itself, allowing AI to complete high-frequency transactions at millisecond speeds.

In traditional on-chain payments, the interaction itself is slow and complex. x402's approach is to have the Facilitator act as an "execution agent" for on-chain transactions. It is responsible for verifying signatures, advances gas, submits transactions, and handles on-chain details, while the payer only needs to submit a signature to the Facilitator, without needing to directly complete on-chain operations. For both buyers and sellers, this greatly simplifies the payment process because the Facilitator solves the trust and settlement issues for them.

Q: What are your views on the development prospects of the x402? What problems or limitations might it face in its actual implementation?

Yang Haipo: In terms of future prospects, I believe the value of x402 lies primarily in the agent-to-agent economic network, rather than in the end-user payment experience. For ordinary users, payment should be completely seamless. In the future, you won't see AI agents prompting you to "scan to pay." When you give the instruction "Analyze market trends for me every morning at 9 AM," the agent will automatically access news or social media data from multiple service providers in the background. Regarding potential costs associated with high-frequency calls, the agent can autonomously pay and obtain services through the x402 protocol, without any human intervention. This model will drive API calls from the traditional "subscription system" to true "pay-as-you-go," because x402 is naturally suited to this type of high-frequency, fragmented collaboration between machines.

Furthermore, there's a security advantage that's easily overlooked. Today, if you wanted AI to buy things for you, you'd almost never dare give it your Visa card number, because credit cards inherently involve unlimited liability. If the agent is attacked or experiences hallucinations, it could indeed "max out your card." However, a token wallet allows you to set an authorized limit for the AI, such as a 100 USDC "spending money account." This way, even if the agent malfunctions, the losses are manageable.

However, precisely because x402 is designed so simply, its shortcomings are also very obvious. First, the x402 protocol is heavily reliant on facilitators like Coinbase. While this simplifies development, it introduces a single point of failure. If a facilitator's server goes down, or if it acts maliciously or censors your transaction, the entire payment chain breaks down—a classic example of centralized risk. On the other hand, x402's very simple design also results in certain functional deficiencies, such as "refunds." Currently, the x402 protocol does not have a built-in refund mechanism. Real-world commercial payments require handling numerous disputes, such as incomplete services or damaged goods, and the irreversible nature of x402 makes these processes difficult to implement.

Against this backdrop, the industry is also exploring broader and more universal agent payment protocols, such as Google's AP2. They are attempting to establish a unified standard that can accommodate traditional Visa/Mastercard, support cryptocurrencies, and take into account complex business processes like refunds. In the long run, a large-scale protocol like AP2 is certainly the direction we hope to see, but precisely because of its complex design and the large number of stakeholders involved, it is still far from being truly implemented. x402, on the other hand, excels in simplicity. It doesn't require waiting for banks to upgrade their systems; as long as you have a wallet and some code, it can be used today.

Q: Returning to the present, from your personal observation, in which areas are AI agents mainly being implemented and generating real value?

Yang Haipo: Frankly speaking, the biggest beneficiaries of AI agents right now are actually the developers themselves. AI pair programming has become a daily routine for many engineers, and agents like Cursor have been widely adopted. For large, complex projects, developers certainly won't entrust the entire responsibility to AI, as that's unrealistic at this stage. However, for some tedious but time-consuming tasks, such as code review, unit testing, and even some algorithm logic generation, AI agents can already handle a significant portion of the workload, greatly saving developers' time.

Another noteworthy scenario is its assistance to non-technical personnel. For example, the recently popular "Vibe Coding" has gained popularity because it unlocks the imagination of non-technical individuals. Previously, without programming skills, even with good ideas, you couldn't implement them. But now, you can use natural language to tell your ideas to the Agent, which will then write the code for you. Of course, we must also be realistic; Vibe Coding is not a "one-click" all-purpose tool, and the Agent's output often requires repeated debugging. Furthermore, while it enables rapid deployment, after multiple iterations, the code can easily become bloated and messy, posing significant challenges for subsequent maintenance and upgrades. Nevertheless, even with a current success rate of only 30-40%, the breakthrough it provides for non-technical personnel—from zero to one—remains extremely valuable.

There are also some seemingly small but common needs in actual work. For example, developers occasionally need an icon, a button style, or a simple interface sketch. In the past, they could only ask their design colleagues for help. Now, the Agent can help you quickly generate an icon or a first draft UI, even if it's just a "usable draft," which can save a lot of time spent on back-and-forth communication.

Although AI is not perfect in many ways, it is good enough for small teams and independent developers who want to create a demo or MVP.

Q: From your observations, what are the future potential possibilities for AI agents? Is it possible that similar new attempts will emerge in the crypto industry in the future?

Yang Haipo: If we look at it from a longer timeframe, I think the potential of AI Agents is definitely not limited to the current development and assistance level. There are many possibilities for the future, such as more autonomous collaboration and autonomous procurement tasks.

There are already some interesting explorations in the industry. For example, Nof1's AI live trading competition essentially allows agents under different models to test their strategic capabilities in a real market environment. In this case, AI no longer just provides information, but forms its own closed-loop action.

Furthermore, an increasing number of exchanges are beginning to support MCP (Editor's note: Model Context Protocol). For example, CoinEx, an exchange within our ecosystem, has already released its corresponding MCP service on GitHub. With the help of such an MCP service, the AI Agent can directly access real-time market data, candlestick charts, news feeds, and other information from the exchange, and then perform in-depth analysis in conjunction with the model. Theoretically, it can not only automatically generate strategies based on user preferences and risk parameters, but if deployed locally, it can also automatically place orders. In this scenario, the AI Agent will truly possess capabilities such as automated trading and intelligent market making. For example, it can dynamically adjust order prices and quantities by acquiring real-time data on market depth, volatility, and trading volume, thereby improving market efficiency and liquidity. The emergence of such capabilities signifies that the Agent has moved from simply "helping you find information" to "helping you make decisions and execute them."

In this model, we can also see the application of x402. For example, if you ask your Agent A to write an in-depth Bitcoin analysis report, but it doesn't have the relevant data itself, it will automatically call other Agents to complete the entire chain: requesting on-chain holdings and transaction data from Agent B, which is responsible for on-chain data monitoring, and automatically making a payment; requesting a sentiment summary from Agent C, which is responsible for news aggregation, and making another small payment. Although the result you see is just "receiving a report," multiple Agent-to-Agent microtransactions have actually occurred behind the scenes.

From the existing examples, we can see that Nof1 proves AI can make decisions, MCP solves how AI acquires data and executes, and x402 enables AI to collaborate economically with other agents. Therefore, I believe the potential of AI agents lies in two directions: stronger autonomous decision-making and more natural economic collaboration. When agents can automatically find resources, purchase services, call tools, and complete the entire task chain, what we see will no longer be the capabilities of a single model, but a digital economic system composed of multiple agents. These things may only be in their early stages today, but I believe the future of AI agents is full of possibilities.