From AI Assistants to On-Chain Agents: What's the Next Evolution of Web3 Wallets?

Over the past year, AI+Crypto has been one of the most dominant narratives in the cryptocurrency industry.

Initially, the market was more focused on AI tokens, decentralized computing power, data networks, and various AI Agent concept coins. However, entering 2026, projects are no longer just discussing "how AI will change Crypto," but are beginning to embed AI into wallets, exchanges, payment protocols, and on-chain execution processes.

Within just one week, three landmark events occurred in succession:

• On April 24, Binance Wallet launched Agentic Wallet, an independent, non-custodial wallet purpose-built for AI Agents. It features an independent balance, configurable permissions, real-time monitoring capabilities, and comes with built-in Agentic Wallet Skills, enabling Agents to perform automated operations within predefined boundaries;
• On April 28, TON launched Agentic Wallet, a self-custodial wallet contract for AI Agents. Users can set up an independent wallet for an AI Agent in minutes, allowing the Agent to autonomously execute on-chain operations such as transfers, Swaps, and staking without touching the user's private keys;
• On April 29, OKX released the Agent Payments Protocol, an open payment standard designed for AI Agent business activities. It covers the complete commercial workflow including quoting, negotiation, escrow, metering, settlement, and dispute resolution, attempting to provide a practical payment and settlement framework for the future Agent economy;

On the surface, this looks like a giant race centered around "letting AI take over on-chain execution." But if you zoom out, the real underlying themes of this competition are permissions, reuse, and co-creation.

This might precisely be the most significant change for Web3 wallets in the next decade.

1. The Evolution of AI+Crypto

Let's first review the evolution of AI+Crypto over the past few years.

As is well known, for most ordinary users, the biggest problem with the on-chain world has never been a lack of information, but rather too much and too messy information. For example, a single token might simultaneously appear on 𝕏, Telegram, DEX Screener, block explorers, project documentation, and various KOL discussions, making it extremely difficult for most users to judge where the risks and opportunities lie within a limited timeframe.

Therefore, the first AI functions implemented by many wallets and trading platforms revolved around "information understanding."

Several AI features launched by Binance Wallet in January this year are typical examples: Social Hype attempts to rank token popularity using social attention and interaction data; Topic Rush generates topic cards for emerging narratives and categorizes them by capital inflow stages; AI Assistant provides summaries of narratives, sentiment, event timelines, etc., on token pages.

The value of these features isn't to make decisions for users, but to help lower the barrier to understanding. For on-chain beginners, a clear AI summary might be more helpful than dozens of fragmented tweets. Even for experienced users, AI can act as an information filter, helping them quickly locate leads worth further investigation.

But this is only the first step. In the Crypto world, after "understanding," users often need to consider "how to operate" – things like authorization, cross-chain bridging, setting slippage, managing risk, revoking permissions, and identifying phishing pages. In other words, if AI remains only at the level of "helping users summarize information," it still cannot truly enter the core process of on-chain interaction.

So, if the first phase of AI was more like a research assistant, the second phase of AI is moving towards being an executor (for further reading: New Tickets for the AI Agent Era: Pushing ERC-8004, What is Ethereum Betting On?).

Gemini's Agentic Trading is a very representative signal. According to Gemini, users can connect an AI Agent to their trading account via MCP, allowing the Agent to call trading-related APIs to perform actions like reading market data, placing orders, monitoring markets, and managing risk. Gemini has also designed certain trading capabilities as Trading Skills, such as getting real-time quotes, checking bid-ask spreads, and reading candlestick data, enabling the Agent to call these modules like tools.

This indicates that exchanges are rethinking the future interaction between users and trading systems. It may no longer be about clicking buttons, entering prices, and confirming orders, but rather users setting goals in natural language, with AI assisting in breaking down the path, calling tools, and executing.

Similar changes are occurring at the wallet layer.

Cobo's Agentic Wallet proposed the Pact protocol to define what an Agent can and cannot do, and under what conditions it should stop. Simultaneously, it uses a Recipe-driven skill layer to provide Agents with predefined execution paths, such as lending, Swap, DCA, grid strategies, and micropayments. Coinbase is also developing Agentic Wallets and Agent Skills for AI Agents, combining them with machine payment protocols like x402, allowing Agents to send, trade, earn yields within set limits, and even automatically pay for resources like APIs, computing power, data, and storage.

The core change behind all this is that Crypto is providing AI with a set of financial and account infrastructure that can be directly invoked. Looking at these actions on a single diagram reveals a high degree of structural convergence – everyone is assembling the same four pieces of the puzzle:

• Identity Layer: The on-chain identity and recognizability of the Agent (OKX Agentic Wallet, Coinbase Agent identity);
• Permission Layer: The granularity and duration of user authorization for the Agent (MetaMask ERC-7715 advanced permissions, Coinbase safety guardrails);
• Payment Layer: Micropayments, subscriptions, and settlements for Agents (x402, APP, MPP);
• Execution Layer: Agents placing orders, cross-chaining, and calling contracts on exchanges/public chains (Gemini Agentic Trading, TON Agentic Wallet);

After all, in the traditional internet, while an AI Agent can help you search, summarize, write emails, and generate code, it encounters identity, trust, account, and settlement issues once it enters payments, authorizations, and asset operations. Blockchain, on the other hand, natively provides accounts, signatures, assets, payments, contracts, and traceable records, giving AI Agents the opportunity to evolve from "digital assistants" to "on-chain executors."

However, execution capability is just the first step.

If these infrastructures are about "building hands" for the Agent, then another more important, and often overlooked, question is: where do the Agent's "working methods" come from?

2. AI Agents Need More Than Just Wallets; They Need Reusable Skills

This is why the Skill concept is worth paying attention to.

How to understand Skill? Essentially, it is a capability unit that encapsulates specific working methods. For example, each Skill is like a folder containing a SKILL.md file that tells the AI what the Skill does, when to use it, and how to operate it specifically.

When runtime, a large model like Claude only needs to load the corresponding Skill on demand based on user intent, transforming from a general-purpose model into an Agent specialized in a specific task.

While this design seems simple and even somewhat "anti-high-tech," the truly important aspect behind it is that the emergence of Skills means competition in the Agent ecosystem is no longer just about model parameters, reasoning capabilities, and invocation speed, but also about working methods, community knowledge, and reusable capabilities.

The moat of an Agent platform might no longer be "my model is bigger, faster, and smarter," but rather "on my platform, can experts, enterprises, developers, and ordinary users encapsulate their experience into reusable assets that can be repeatedly invoked by others and other Agents."

Applying this judgment to the wallet industry yields an even more interesting proposition: As wallets enter the Agent era, the competition in the second half is not just about executing well, but about becoming a workspace where users can沉淀 methods and co-create.

In fact, whether it's Gemini's Trading Skills, Coinbase's Agent Skills, or Cobo's Recipe-driven skill layer, they are all essentially doing the same thing: breaking down complex operations into standardized, modular, and invokable capability units.

This is particularly important for ordinary users, as on-chain operations are rarely simple instructions. An seemingly ordinary Swap might involve token identification, contract address verification, allowance management, slippage setting, Gas estimation, phishing risk identification, and transaction confirmation. A cross-chain operation might involve bridge selection, arrival time, fees, target chain asset display, and failure handling. Participating in DeFi involves understanding yield sources, smart contract risks, liquidation risks, and exit paths.

If this experience exists only in someone's mind, it is difficult to transfer. If it's just written as a regular tutorial, it's hard to be automatically invoked in real operations. But if they can be distilled into Skills, templates, checklists, or interactive works, they can become knowledge assets reusable by every user in the AI era.

This is the new value of wallets in the AI era.

Looking ahead, what truly matters in the future wallet track is not just "can AI help me complete an operation," but whether the experience behind these operations can be沉淀, reused, and collectively improved.

This is precisely the direction imToken can further explore when introducing UI 3.0.

3. The Next Decade: A Vision for Wallets in the AI Era

Objectively, over the past decade, the wallet industry has undergone multiple iterations: from mnemonic phrases to multi-chain asset management, from DeFi to NFTs, from L2s to account abstraction, and now to AI Agents.

Regardless of how the industry narrative changes, wallets have always faced the same fundamental question: how to enable users to manage their on-chain life more securely, autonomously, and clearly. In the AI era, this question becomes more complex and more important.

Because when users start involving AI in on-chain operations, wallets need to answer not just "can this transaction be signed," but also: what does this transaction mean? Does it align with the user's true intent? Is the authorization scope reasonable? Have the risks been adequately explained? Especially if the Agent makes a mistake, does the user still retain ultimate control? If this operation has value, can the user 沉淀 the experience and share it with newcomers?

This is also the new direction UI 3.0 can convey. The future wallet interface isn't just about looking better or feeling smoother, but about understanding users better, explaining risks better, carrying knowledge better, and being more suitable for user co-creation.

From this perspective, the considerations imToken has made around Verifiable UI over the past year form the first baseline for wallets in the AI era (for further reading: From the Kelp DAO Incident to Verifiable UI: Why 'Verifiable Interface' Could Be the New Decentralized Security Baseline?).

This is a crucial, somewhat counter-intuitive judgment. In most industry discussions, AI often implies more natural input, smarter parsing, and smoother human-computer dialogue. But from a wallet perspective, the smarter the AI, the more the interface needs to be verified.

What Verifiable UI aims to solve is precisely that users must have the ability to verify the authenticity of the interface itself. Even if centralized services go offline, or if the Agent's interpretation deviates, users should still retain ultimate control over their assets and interaction processes. This is homologous to the non-custodial principle imToken has adhered to for years.

Non-custodial addresses the "who holds the private key" issue, while Verifiable UI further addresses the "is what I see exactly what I am about to sign" issue.

It is within this baseline that imToken believes "constraining probability with certainty" is more suitable for wallet judgment in the AI era. After all, no matter how smart the AI or how powerful the model, whether a final transaction can be executed must still be subject to the deterministic code, rules, and permission boundaries within the wallet. The wallet is not an ancillary entry point for AI, but the ultimate checkpoint for user asset security in the AI era.

But beyond defense, UI 3.0 has another more ambitious main line: transitioning from a tool to a co-creation platform. When execution is gradually taken over by Agents, and the interface becomes verifiable, the next question wallets must answer becomes: what can users create here? The logic behind this is isomorphic with the logic of Skill.

As more and more wallets integrate Agents, design permissions, and provide execution capabilities, long-term differentiation may not come from "who integrates AI first," but from whether users can 沉淀 their own works within that wallet. This could be an on-chain strategy, a security rule template, a wallet application for a specific scenario, a shareable operation checklist, or a Skill that can be invoked by other users and Agents in the future.

Overall, the evolution of wallets over the past decade has been from "address book + signer" to "DApp browser" to "multi-chain entry point." In the next decade, the question they need to answer might be: when Agents execute on my behalf, and the interface is verified by me, what can I still create together here?

Because when users start involving AI in on-chain operations, wallets need to answer not just "can this transaction be signed," but also: what does this transaction mean? Does it align with the user's true intent? Is the authorization scope reasonable? Have the risks been adequately explained? Can the user 沉淀 the experience of this operation and share it with newcomers?

This is also the new direction UI 3.0 can convey: the future wallet interface isn't just about looking better or feeling smoother, but about understanding users better, explaining risks better, carrying knowledge better, and being more suitable for user co-creation.

This means users aren't just sharing stories; they are also 沉淀ing works together. For example, a Skill on "how to safely add a token," an interactive guide for "using Layer2 for the first time," a risk checklist for "identifying phishing authorizations," a process template for "hardware wallet cold storage," or even a learning path for beginners.

These works don't have to be complex initially. They can be an image, a tutorial, an operation checklist, a security reminder template, or even a Skill prototype that can be invoked by AI Agents in the future.

The important thing is that users transform from "being educated" to "co-creating."

This might be the most anticipated change in user education in the AI era: knowledge is no longer unilaterally output by projects, but is continuously supplemented, corrected, verified, and reused by real users in real scenarios.

In Conclusion

Looking back, the integration of AI+Crypto is rapidly moving from concept to product.

Exchanges are starting to connect AI to accounts and trading capabilities. Wallets are designing asset permissions and execution boundaries for Agents. Payment protocols are exploring automatic settlement between machines. Skills and Recipes are beginning to turn complex operations into reusable modules.

Ultimately, these changes may not point to a world where "everything is done by AI," but rather a world where user capabilities are amplified. In this world, the wallet remains the entry point, but it is no longer just an asset gateway; it becomes the gateway for identity, authorization, knowledge, operations, and co-creation.

For imToken, its tenth anniversary is not just a time to look back at the on-chain journey accompanying users over the past decade, but can also be an invitation to the future. An invitation to every user to 沉淀 their experiences, questions, methods, and works, and to participate together in building the knowledge network for wallets in the AI era.

After all, a truly vibrant Agent ecosystem competes not just on models, but on the community's ability to co-create.

And the wallet might just be the most natural starting point for this community, the path we will all walk together for the next decade.