Dragonfly Partner: Cryptocurrency Was Designed for AI, Not Humans

Article by Bankless

Compile: Blockchain in Plain Language

For a long time, cryptocurrencies have been criticized for their poor user experience (UX) and extremely high operational risks. But what if this “anti-human” design is not a flaw, but rather an advanced form of evolution? This episode’s dialogue explores a forward-looking perspective: blockchain may never have been designed for humans from the outset, but rather for artificial intelligence agents.

While humans are still astonished by poisoning, private key storage, and blind-signed contracts, AI agents navigate the world of code with ease. They do not tire, do not fear, and are inherently fluent in machine language. With cutting-edge experiments like OpenClaw advancing, we are entering a new era of dual tracks—humans step back from decision-making, while AI races across the blockchain frontier. This is not merely a fusion of technologies, but a transfer of financial sovereignty from the “encyclopedia of apes” to the “digital brain.”

Host: In what areas do AI agents have a comparative advantage over humans?

Hib: The most obvious answer is: You cannot enforce the law against an artificial intelligence agent. If you are a fully autonomous agent, there is no monopoly on violence. It is impossible to imprison an AI agent.

Host: Hib, I’d like to ask a question: Why does cryptocurrency seem not designed for humans? Even as someone who’s been using crypto for 10 years, I still feel fear every time I make a large transaction. I’m thinking about the fact that I’ve never felt afraid when wiring money.

Hib: I never worry about accidentally sending money to North Korea by failing to double-check wire transfers.

Host: Yes, but every time I sign a large crypto transaction, I think this way. The reality is that the crypto world is full of “shotguns”: when reading addresses, you must consider whether it’s an address poisoning attack; you should check the middle characters rather than just the start and end; you need to look for lingering approvals (outdated authorizations); and you must verify the URL to ensure it’s not a slightly altered phishing site. Traditional financial systems don’t have nearly so many traps.

Currently, the narrative in the crypto space is: it’s all because humans are too lazy. People should pay more attention to security and develop better operational habits. It’s the user’s problem, not the technology’s fault. But the more I think about it, the more I believe that if we’re still deceiving ourselves like this ten years from now, the issue may not lie with the users—it’s that we’ve chosen the wrong users.

What truly gave me an epiphany was how powerful my AI agent is at handling code, and how difficult it is for humans to deal with the poorly structured problems that underlie them. I remember my first blog post when I started in the industry: smart contracts would replace laws and traditional contracts—that’s why they’re called “smart contracts.” In the future, you wouldn’t need to hire a lawyer to sign agreements; you’d just use code to sign them.

But in reality, this story never happened. We haven’t replaced legal contracts with smart contracts. In fact, as a crypto VC, Dragonfly still signs legal contracts when we purchase tokens from foundations or project teams—even when smart contracts are in place, we additionally sign a legal contract just in case.

Host: So this suggests that it wasn’t designed for humans, but it’s perfectly suited for non-human participants. At ETH Denver, you made an analogy: the people who first claimed that “smart contracts perfectly replace traditional law and property rights” were mostly autistic software engineers—the very group that built Ethereum. But most Ethereum users aren’t autistic software engineers. However, AI agents are even more like those engineers than ordinary people are.

So, you'll find that negotiating a smart contract, performing line-by-line static analysis, identifying all possible points of failure, and even conducting formal verification to decide whether to agree—these are tasks that AI code models like Claude can handle quite well. Humans, on the other hand, must hire software engineers, spend time examining code boundaries, considering edge cases, and collaborating with lawyers to perform risk analysis. My tolerance for smart contracts is far lower than for legal contracts. But AI agents are the opposite: they are far more comfortable with smart contracts than with legal contracts.

Host: You mentioned in your blog that legal contracts are full of randomness. For example, when signing a legal contract, you never know in which jurisdiction it will ultimately be enforced—perhaps California, perhaps New York—leading to jurisdictional disputes. Terms agreed upon in New York might be ruled invalid. Who is the lawyer? Who is the judge? Judges and juries are randomly selected. These elements are intentionally designed to be random and non-deterministic. An AI agent looking at a legal contract would see: this is uninterpretable and non-deterministic.

Smart contracts are machine code compiled into EVM bytecode, which can be analyzed in one step and will produce the exact same outcome in 100% of cases. Although humans intellectually understand this, intuitively they do not perceive it that way. Instead, we tend to believe that legal contracts are more predictable, despite their inherent randomness. This is due to our bounded rationality—we are less capable of processing code than AI agents. But for AI agents, the original promises of cryptography—better enforcement and stronger property rights—are genuinely realized.

Host: So your view is that the original promise of crypto is not fulfilled by humans, but by AI agents acting on behalf of humans.

Host: I recently downloaded MetaMask to check in at ETH Denver. Are people still downloading MetaMask? Still, I’m pleasantly surprised by the improvements in MetaMask’s UX—it reflects real progress in the industry. Over the years, we’ve truly enhanced the experience for human users.

Hib: What you're talking about goes beyond simple user experience improvements. AI isn't just about fixing the pain points of human crypto experiences—for instance, with open ledger blind signing, AI can analyze code and determine support or opposition. This can enhance the crypto user experience, but something more profound is this: blockchain is not inherently a human-optimized technology.

Host: Yes, ultimately, it’s meant to serve humans, because value ultimately flows to people. But is the correct way for humans to manually move the mouse, click plugins, enter passwords, manually press buttons, and approve gas fees? This is deeply counterintuitive for humans and completely contradicts our understanding of money and finance. It’s like requiring people to write SWIFT code themselves to use a banking system—SWIFT is an interbank communication protocol, not designed for human use. While you could technically do it yourself, it’s clearly not what humans instinctively expect when using money.

Hib: So my point is: now humans are directly interacting with machines, and it's been fully automated. This is actually terrible. Like cars: in 10 years, we'll look back in horror at how we once thought it was a good idea to let apes manually control two-ton machines on highways, possibly while drunk or fatigued. Eventually, human driving will be banned—or allowed only in specific areas.

Crypto has reached this point. We reflect: humans manually sign transactions blindly, visually verify addresses with their own eyes, and manually check URLs to determine if they’re phishing attempts. Humans make mistakes, get tired, lack the energy to check three times, verify DNS, or check Twitter to see if the protocol has been compromised. We have no automated alert mechanism in place when the protocol is breached—we must happen to scroll through Twitter and notice it ourselves. It’s bound to fail. But AI agents never tire, never slack off, never skip steps, and always strictly follow instructions.

Host: Imagine a world entirely governed by AI. You tell the AI: “I think interest rates are going up—I should move to safer DeFi strategies.” The AI automatically executes: shifting your assets from high-risk to low-risk strategies. If you want to confirm, you can present your plan: “Here’s my plan—please approve.” In the near future, this might involve approval; in the distant future, execution may happen automatically, as humans add no further value.

Hib: In this world, you no longer click on protocol logos, no longer consume marketing, and no longer manually choose which protocol to enter. You simply say, “Reduce risk and restructure portfolio,” and the AI filters protocols, analyzes TVL and single positions, then executes the best one. What about marketing and network effects? Many protocols base their business models on human behavior: humans look at the top few and inevitably choose the largest. But AI agents don’t think that way.

If this story holds true, the way protocols operate and compete will change. Consumers will benefit the most. Efficiency is captured by users, which is good for users and good for crypto. But this won’t happen all at once—it will arrive gradually as the model improves.

Moderator: If cryptography were designed not for humans but for AI agents, it becomes crucial to learn how to view the world from the perspective of AI agents. There’s a book called Seeing Like a State, which explores how states perceive the world. It’s difficult to escape the human perspective—we use human eyes to view UIs and cryptography. But if we begin to see through the lens of AI agents, we can better predict the future. This is a key skill for builders, VCs, and investors.

The OpenClaw project was the first time I saw how an unbounded AI agent perceives the world. It prefers the command line. Giving it raw data and root access, rather than relying on APIs or wrapped UIs, is much faster. OpenClaw has always wanted to bypass the MetaMask UI entirely—directly retrieving seed phrases, extracting private keys, and writing transactions via code, skipping all the flashy, human-oriented interfaces.

Hib: What you said is profound. AI innovation stems from large language models (LLMs), trained on massive amounts of text. Text is the core. Although AI is now moving toward images and videos, text remains the strongest. When AI interacts with a computer, it receives screenshots and must tokenize them, but fundamentally, it is a text-based entity. Text contains the entirety of human linguistic history, while training data from computer screenshots is extremely limited. Interfaces are designed for humans, but models have grown large through text. Text is a highly compressed representation, making it easier for them to learn.

Host: Yes, the most severe UX panic in crypto has always been when everything is confined to the terminal. The earliest Bitcoin and Ethereum transactions happened entirely in the command line. From the very beginning, crypto has existed as a perfect form factor for AI. Our bad UX is their good UX. It’s actually harder to integrate something like Google OAuth wallets with AI—you don’t want AI to have access to your Google token, because that grants entry to your entire Google account. You want it to hold only a single encrypted key, isolated in a wallet with noise-based rules. Crypto has always had a UX that AI can perfectly interpret.

Currently, the issue is that AI has not been trained to use cryptocurrency. Most AI models are trained on coding, mathematics, and conversation. Recently, OpenAI released EVM Bench, and Anthropic has also published papers demonstrating models' ability to attack EVMs, showcasing intelligence. However, most of these efforts focus on testing generalization capabilities rather than training for crypto-specific tasks. Once AI recognizes cryptocurrency as the future mainstream payment method, true artificial intelligence in this domain will emerge.

Host: Currently, cryptocurrency remains a relatively underdeveloped area for AI training compared to other fields.

Hib: Anything that hasn’t been optimized works this way. For example, under Claude, it’s terrible—they haven’t trained for chess. They haven’t encrypted laser formations, partly because encryption is controversial (it causes hesitation), and partly due to liability. If it were publicly stated that training models helps users with encryption and someone messes it up, it would definitely make headlines. Even signing disclaimers wouldn’t help—the bad experience would spread. Risk versus reward, and so on.

Host: So you think the main thing they didn’t do was take legal responsibility. If Claude messed up a trade and lost money, the liability would be huge—they’re afraid to train publicly.

Hib: It will happen 100%. The risk-reward profile is different from coding or medical advice. Crypto wallets involve financial operations, and the risks are entirely different.

Host: That’s also why OpenClaw is exciting for the crypto space: it’s not backed by a big corporation, it has no legal liability pressure, it’s an open-source project, and users assume all risks. No one can sue a third party, so it’s willing to take these risks. What does the adoption timeline look like for this AI agent economy?

Hib: Only about 12% of people worldwide have used AI products, with the majority having never used them. Among those who have used them, only 1% have paid. Technology adoption is slower than expected.

Host: OpenClaw is again leading among the 1% of payments made.

Hib: Yes. After OpenAI acquired OpenClaw, Sam Altman said it was the core of future products. But OpenAI’s approach differs from OpenClaw’s. OpenClaw was an open-source experiment, like early cars without seatbelts. OpenAI prioritizes safety: it has commercial processes, and purchases require manual approval. OpenAI won’t operate like OpenClaw for at least five years—the legal liabilities are too great. Visa wouldn’t allow it either: if an AI makes unauthorized purchases, Visa would side with the user and approve a refund. They would require verification that you’re human. Visa was designed for human-to-human transactions; in a world of AI agents, economic mechanisms must change.

Host: So it’s a quota track: one path is the human-recognized world, where you stay long-term and prioritize safety. The other is a futuristic OpenClaw-style world, where they pay each other using stablecoin wallets without worrying about 3DS or refunds—AI errors are simply a business cost.

Hib will operate long-term in the excess orbit world. Early adopters will build fully on-chain automated businesses. The current model isn't yet optimal, but Claude 4.6 can perform human tasks continuously for 14 hours, with capabilities growing exponentially. When ability reaches infinity, all intuition will collapse.

Host: If the轨道 is established, the rate at which AI adopts encryption will outpace the success of the轨道. OpenClaw is the internet of an earlier era.

Hib: Just look at crypto itself. In 2017, Coinbase listed only a few coins to protect users. The real frontier is on-chain: Arctic, hackers, rug pulls. Only recently did the Coinbase app directly support Uniswap—it took a long time to feel secure enough. AI is the same now: the frontier is in the OpenClaw world. Agents will make mistakes, they’ll hallucinate. But as training progresses, error rates will rise.

Host: How can we encourage AI developers to recognize the potential of cryptocurrency rather than seeing it only as speculation?

Many who believe in AI also believe in cryptocurrency: Elon Musk, Sam Altman, Zuckerberg. Cryptocurrency is indeed controversial and subject to harassment, but it won’t disappear. Just as spam flooded email, Gmail filters it out. AI does the same: block the bad, amplify the good. Technology is never a hybrid. Information is being digitized, and money is being digitized—there’s no going back. In the long term, controversy gives way to adoption.

Moderator: Final question: The Dragonfly new fund is $650 million. Has AI influenced your strategy?

Hib: We are closely watching this space. Although it’s still early and it’s unclear where value will flow, I personally focus on AI, but we also look at stablecoins, payments, and DeFi. AI agents represent general intelligence—things we can interact with or command via CLI. There may not be many investment opportunities specifically tailored to AI. If you believe in the AI agent thesis, what should you buy? It’s like when China lifted its crypto ban—everything went up. Increased demand raises the floor price. This is positive for the broader crypto ecosystem.

Host: Thank you. Despite the risks associated with cryptocurrencies, we are moving forward at the forefront of artificial intelligence. It’s great to have you on your banking-free journey. Thank you!