Buterin: AI Agents Replace Wallets, Talk Directly to SDKs

Vitalik Buterin is proposing to delete the single most-invested piece of consumer crypto: the wallet interface. In a recent interview, the Ethereum co-founder framed AI agents as the replacement for UI across the entire stack, with humans demoted to principals who never touch a "Connect Wallet" button again. For an industry that has spent roughly a decade optimizing MetaMask-style flows, seed phrase onboarding, and WalletConnect handshakes, that is a call to write off a lot of accumulated UX work.

The claim matters because Buterin isn't arguing for a better wallet. He is arguing that the wallet, as a visual artifact, goes away, and that on-chain infrastructure becomes the coordination substrate for AI agents built by different teams that don't trust each other. That reframes Ethereum's product-market fit from "world computer for humans" to "settlement layer for machines".

Key Details

As CoinMarketCap reported, Buterin split his argument into four connected pieces: the role of humans, the role of Ethereum, the shape of wallets, and the cryptographic primitives that make agent-to-agent interaction workable.

On humans, he was explicit: "I personally still think of people as being the users. I think of AI as being the replacement for UI." Agents act, humans remain the principals whose interests are being served. That is a specific ontological choice, not a throwaway line. It rules out the fully autonomous-agent-as-economic-actor framing that some AI x crypto projects have been marketing.

On Ethereum itself, Buterin described the network as two things today: a public data layer where anyone can post information on-chain, and a computation layer for decentralized applications. Both, he argued, stay relevant in an AI-driven environment. What changes is the interaction pattern: agents managing simultaneous operations across wallets, protocols, and external services on behalf of users.

On wallets, the prediction is blunt. Traditional wallet interfaces could disappear entirely. AI systems would interact directly with wallet SDKs and APIs. "Maybe you don't need a wallet. You just talk to the SDK." He also noted that agent workflows may tolerate slower execution, particularly during multi-step reasoning, which is a meaningful shift from the sub-second UX bar wallets have chased.

On cryptography, zero-knowledge proofs sit at the center. Buterin wants agents and users to verify credentials, reputation, and fund sources without revealing full identity. "You can make zero-knowledge proofs about your reputation. You can make zero-knowledge proofs about the source of your funds." The design principle is selective disclosure: minimum information per interaction.

He also flagged a governance risk: "A lot of things that used to require a high amount of effort to break, you can break them automatically." On layer-2, he criticized the current pattern of copying the EVM, scaling it up, adding centralization, then anchoring back to L1, and argued future L2s should be shaped around specific applications like privacy, trading, or specialized execution. ZK payments, in his view, are the likely next step for native AI agent payment standards, and the Ethereum team is building a ZK API that separates each agent request from every other without requiring one on-chain transaction per action, combinable with security deposits to deter abuse.

Why This Matters for Crypto and DeFi

Strip the vision down and there are three architectural bets stacked on top of each other. First, that agents become the dominant transaction originators. Second, that no single AI vendor gets to be the trust root, which forces settlement onto neutral infrastructure. Third, that the privacy layer has to be cryptographic rather than policy-based, because agents will interact millions of times per user and any identity leakage compounds.

If bet one holds, wallet vendors are in trouble. The moat of a MetaMask or a Rabby is largely the front-end, the extension distribution, and the user's muscle memory. If the "user" is a language model calling an SDK, none of that is defensible. What survives is key management, policy engines, and transaction simulation, which are backend concerns. I'd argue this pushes value toward embedded wallet infrastructure providers and account abstraction stacks that already expose programmable APIs, and away from consumer-brand wallets.

Bet two is where Ethereum's pitch sharpens. Buterin's framing that on-chain infrastructure is "the only viable trust and settlement layer" for agents built by different developers is a direct answer to the question of why crypto matters in an AI world. It is also unfalsifiable in the short term, which is worth flagging. The source doesn't disclose what volume of inter-agent interaction actually needs a blockchain versus what can be done with signed receipts, escrow services, or reputation networks off-chain. That distinction matters because it sets the bound on how much of the AI economy Ethereum can plausibly capture.

Bet three, ZK everywhere, is the one with the most technical debt to pay down. The proposed ZK API that unlinks agent requests without one transaction per action is architecturally interesting, but the source does not specify proving costs, latency, or the trust assumptions of the aggregation layer. Those are the numbers that decide whether this works for high-frequency agent payments or only for coarse-grained settlement. Standards work at the EIP level will tell us which direction this goes.

Industry Impact

For engineering teams building in DeFi, iGaming, and fintech-adjacent crypto, the near-term implication is that API surface area beats UI polish. If agents become the primary caller, your protocol needs machine-readable intent specifications, deterministic error semantics, and rate-limit behavior that a planning loop can reason about. Human-friendly error messages ("insufficient gas, please try again") are worthless to an agent that needs a structured code and a suggested remediation.

Account abstraction and session keys become table stakes rather than optional. An agent operating across dozens of protocols cannot pop a confirmation dialog for every action, so scoped permissions, spending limits, and revocation primitives have to live in the smart contract layer. Teams that have been slow-walking ERC-4337 integration are going to feel that pressure sooner than expected.

For iGaming and prediction market operators, Buterin's governance warning deserves attention. If automated attacks lower the effort barrier, any mechanism that assumed human-scale friction, whether that's captcha-gated bonus abuse or manual review of large withdrawals, needs a rethink. The defensive posture shifts toward cryptographic proofs of personhood, reputation attestations, and on-chain incentive slashing rather than operational checks.

For infrastructure providers, the tolerance for slower execution in agent workflows is a small but real gift. It opens design space for proving systems, batch aggregation, and privacy-preserving computation that would be UX-unacceptable in a human-facing wallet. What we don't know, and the source doesn't quantify, is how much slower is acceptable. If it's 2x current wallet latency, that's a modest opening. If it's 30 seconds for a multi-step reasoning transaction, that's a completely different design envelope. That bound is the single most important unanswered question here, because it determines which cryptographic techniques become viable in production.

What to Watch

Concrete signals to track over the next 12 to 18 months. First, whether the Ethereum team ships the ZK API with published proving cost benchmarks. Without numbers on cost per request and latency, it stays a whiteboard design. Second, whether any major wallet vendor (Rainbow, Rabby, Coinbase Wallet) announces an agent-first SDK or explicitly deprecates parts of its consumer UI. That would be the strongest confirmation that Buterin's read is landing inside the vendor community. Third, whether new L2 launches shift away from generic EVM clones toward the application-specific designs Buterin described. The relevant developer documentation is already public on ethereum.org, and divergence from the standard EVM template should be visible in launch specs.

Testable prediction: if this thesis is correct, we should see at least one top-10 wallet by usage announce an "agent mode" or headless SDK product within 12 months, and ZK-based payment standards should show up as draft EIPs within the same window. If neither happens, the "AI replaces UI" framing is likely running ahead of what the ecosystem is actually prepared to build.

Key Takeaways

• Buterin's core claim: AI agents replace wallet UI entirely, with humans remaining the principals and SDKs becoming the primary interaction surface.
• Ethereum's two roles, public data layer and computation layer, both survive the transition, but interaction patterns shift from single-user clicks to concurrent multi-protocol agent operations.
• Zero-knowledge proofs for reputation, fund source, and per-request unlinkability are positioned as the privacy substrate; the ZK API in development avoids one on-chain tx per action and can be paired with security deposits.
• Governance mechanisms that relied on human-scale attack effort are exposed once AI automates that effort, which forces cryptographic and economic defenses over operational ones.
• Unknown to watch: the source does not specify proving cost, latency, or the acceptable slowdown for agent workflows, and those numbers decide whether the design is a production reality or a research direction.