OKX Wallet Friends | Carnival Edition: A Conversation with Vitalik — The Trends and Changes of Web3 in the AI Era

Previously, OKX Wallet launched OnchainOS and has been iterating to open up Agent capabilities, while also introducing the new Agentic Wallet. As AI becomes the new "user interface," and Agents begin trading and participating in governance autonomously as on-chain participants, how will the role of blockchains like Ethereum change?

At the intersection of AI and Web3, a new interaction paradigm is taking shape.

This is the "Friends of OKX Wallet" series – the Carnival edition. This series, through dialogues with different builders, documents their judgments and reflections at key junctures within the industry. In this edition, OKX Wallet VP Paul Wan will engage in a conversation with Vitalik, exploring the long-term trends and underlying structures of AI and Web3, aiming to understand: in the Agent era, what role will blockchain play.

Evolving Roles of Ethereum and Blockchain in the AI Era

Question 1: Facing the Agent era, what primitives must an on-chain operating system provide? What is still missing in today's Ethereum?

Vitalik:

Here is how I see it. Ethereum primarily provides two major functions for applications. One is a bulletin board, where anyone can publish data on-chain, and various applications interpret this data in many different ways. The other is on-chain computation and off-chain computation, which includes financial applications, DeFi, and various other applications.

In the AI era, fundamentally, these uses are still essentially the same. The same use cases still exist, and these functionalities can continue to remain important. However, AI will definitely cause huge changes in the way we interact with blockchain and other tools.

One key difference is that, in the pre-blockchain era, users interacted through a specific interface, which corresponded to a specific task. In the world of AI, especially the form of AI we see now and in the coming years, you can have a user-side AI that calls upon these different skills, combines things in one go, and interacts with many different objects simultaneously.

This also leads to a massive increase in the number of workflows for interacting with Ethereum and other systems. So from my perspective, the "operating system" analogy is not entirely accurate from certain angles. The operating system will still exist, but it will become smaller and simpler. At the same time, we will have a bunch of different tools and skills that the AI helps users to utilize and accomplish tasks. Blockchain offers a natural choice for enabling multi-party collaborative applications and allowing different participants to cooperate effectively over the long term without needing to establish trust or prior agreement.

Another crucial point is the Economic Layer. If AI becomes more decentralized, there will be many different AI entities, built and controlled by different people, needing to interact with each other. To make this interaction possible, an economic layer is required. Because cooperation fundamentally comes down to two paths: either relying on economic incentives and rules, or relying on centralized control.

If we can build this economic system, it will better enable AIs to interact with each other in a decentralized manner. This is similar to how a complete operating system has a runtime and the software and infrastructure built upon it. In this new economic system centered around Agents, we need mechanisms to discover, define, and match appropriate Agents. Simultaneously, users and their Agents can participate, interact with each other, and build their own Skills, MCPs, CLIs, and strategies.

Question 2: In high-frequency Agent trading scenarios, how should we view L2s?

Vitalik:

L2s are important, but we need to be more imaginative in how we build them. In the past, the approach was often to simply replicate the EVM and scale it, but this method is not ideal. A better approach is to start from application requirements and supplement the capabilities that L1 does not provide. Ideally, different functions should be distributed across different layers: accounts can live on L1, while high-frequency trading and matching can be on L2.

Furthermore, L2s can also take on privacy functions, like Tornado Cash, Railgun, Privacy Pools, etc., which in a sense can be seen as "privacy L2s." We will see more L2 solutions developing in different directions in the future.

Reconfiguring the Relationship Between Humans and Intelligent Agents

Question 3: When Agents can autonomously trade, hold assets, and even participate in governance, how should we redefine the user? Especially from the perspective of on-chain governance, how should we redesign mechanisms for these non-human participants?

Vitalik:

Personally, I still view humans as the users, and I see AI as a replacement for the UI – a new way for humans to interact with the chain.

Think of it this way: before AI, you might need to get information from multiple tools like Google, Wikipedia, and Stack Overflow. Now, you can directly ask an AI, which performs the operations and gives you the result. This change will soon happen with blockchain interactions as well.

This means our perception of infrastructure attributes will change, for example, latency. In human interactions, low latency is usually very important. For Agents, some scenarios require extremely low latency, while for other scenarios, latency may not matter much – a complex problem can wait longer.

This difference will change how we think about the interface layer. Take wallets as an example. What remains unchanged is the SDK, i.e., the API layer (e.g., transferring, querying, privacy operations). But the product form around it might change; users might not even use the wallet directly anymore but interact with the SDK instead.

So we will see a series of carefully crafted software packages with strong security and formal verification capabilities, equipped with Skills files that the AI can call. Therefore, I believe there will be very significant changes in the layer between the blockchain and the user.

Question 4: If the actor of a transaction could be a human, an Agent, or a combination of both, how should we rethink on-chain identity? Can we build a unified framework applicable to humans, Agents, and hybrid scenarios?

Vitalik:

The key to establishing on-chain identity is to decompose identity, proving only the necessary information required to complete a specific interaction. In most cases, fully exposing identity doesn't make sense. A more reasonable approach is to reveal only partial information, for example, proving reputation or the source of funds using zero-knowledge proofs.

At the same time, we need to make on-chain behavior and assets easier to be proven via zero-knowledge proofs, and enable wallets to better help users manage private data. Different applications should adopt different implementation methods, protecting user information as much as possible while meeting requirements.

Building a unified framework applicable to humans and Agents is feasible. Agents can decide for themselves how to distribute tasks across different L2s. Currently, their reasoning methods are not fundamentally different from humans, so the market will gradually find more reasonable allocation methods.

Evolution Paths for Agent Products and Native Standards

Question 5: What constitutes a good Agent product experience?

Vitalik:

Speaking of a good Agent product, it should be intuitive and easy to use while operating as part of a larger ecosystem, not trying to take over the user's entire life. It must also place high importance on privacy and security, an area where many current systems are still lacking.

Ideally, we need more AIs aligned with user interests, not belonging to a single company or application, but acting on behalf of the user. This can reduce attacks and exploitation. Additionally, applications need to be compatible with the user's existing configuration, support personalization, be able to interact with other tools, and ensure security based on this compatibility. If all these can be achieved, it would be ideal.

Question 6: In the Agent era, what is the development direction for wallets?

Vitalik:

On one hand, AI can be used to build Ethereum itself, for example, through formal verification to enhance security, which may even become a necessary capability in the future. On the other hand, it can act as an Agent wallet, integrating various capabilities while ensuring privacy and security.

If this experience relies on third-party servers, true decentralization and privacy cannot be achieved. At the same time, there need to be restrictions on AI behavior, which is also part of the wallet's responsibility in risk control. More importantly, we shouldn't view Ethereum as an isolated system; but rather integrate it into a global AI that helps users accomplish various tasks at the operating system level, including on-chain interactions, internet searches, and local data management, thinking from a full-stack perspective.

Question 7: Under the Agent economy, how should public goods mechanisms evolve? What will the future Native Agent Standard look like?

Vitalik:

Public goods funding is essentially a governance issue, and governance requires defining stakeholders. Behind every Agent is still the person running it, so humans remain central. AI and ZK offer new possibilities for governance, but AI also lowers the cost of attacks, making many mechanisms more susceptible to automated attacks. Therefore, this is a direction requiring continuous exploration and iteration.

Regarding native standards for Agents, there is currently no fully defined form, but an important direction is ZK Payments and ZK APIs. The core goal is: regardless of what type of API request is initiated, each individual request itself is private and completely isolated from every other request.

This is crucial because, in AI scenarios, even if using pseudonyms or anonymous identities, as long as this identity is persistent, with the continuous accumulation of information, re-identification will eventually occur, leading to a loss of privacy. Therefore, it is necessary to ensure, from a mechanistic level, that there is no linkability between individual requests.

The key to achieving this is leveraging zero-knowledge proofs while avoiding putting every single request on-chain; otherwise, both costs and latency become unacceptable. While latency can be tolerated in some scenarios, high costs remain a problem to be solved. On this basis, mechanisms like bonding/staking can be combined to prevent abuse from both the user side and the application side without compromising privacy. There is still a lot of work progressing in this direction.

Conclusion

We thank Vitalik for sharing his thoughts in this dialogue and for providing us with a forward-looking perspective on the intersection of AI and Web3.

Regarding directions like ERC-4337, EIP-7579, and EIP-7702, OKX Wallet has been continuously advancing related exploration and innovation. At the same time, we are closely monitoring the progress of Vitalik's latest proposal, EIP-8141, and look forward to deeper collaboration on the underlying infrastructure in the Agent domain.

Everyone's Web3 moment is different, but the reasons for choosing to stay here are often similar. We thank every friend who was willing to communicate and share with us during the Carnival.

The Carnival has briefly concluded, but the conversation does not end. We will see you at the next moment.