Dragonfly Partner: Cryptocurrency Was Not Designed for Humans—It Was Designed for AI
TechFlow Selected TechFlow Selected
Dragonfly Partner: Cryptocurrency Was Not Designed for Humans—It Was Designed for AI
Blockchain is the machine protocol for AI operation.
Navigating Web3 tides with focused insightsBy Bankless
Translated by Baihua Blockchain
For years, cryptocurrency has been criticized for its poor user experience (UX) and extremely high operational risks. But what if this “anti-human” design isn’t a flaw—rather, a form of prescient evolution? This episode explores a forward-looking thesis: blockchain may never have been designed for humans at all—but instead, built from day one for artificial intelligence agents.
While humans still flinch at phishing attacks, private-key management, and blind-signing contracts, AI agents thrive in the world of code. They don’t tire, don’t fear, and are natively fluent in machine language. With cutting-edge experiments like OpenClaw advancing rapidly, we’re entering a new dual-track era—humans recede from active decision-making, while AI accelerates across the onchain frontier. This is not merely technological convergence; it’s a transfer of financial sovereignty—from the “ape encyclopedia” to the “digital brain.”
Misaligned Users: Why Cryptocurrency Is Inherently “Anti-Human”
Host: In what ways do AI agents hold a comparative advantage over humans?
Hib: The most obvious answer is: you cannot enforce law against an AI agent. If you’re a fully autonomous agent, there’s no monopoly on violence—you can’t throw an AI in jail.
Host: Hib, I’d like to ask: why does cryptocurrency seem fundamentally *not* designed for humans? Even as a 10-year crypto user, I still feel fear every time I sign a large transaction. I’ve been reflecting on this fact: I’ve never feared wire transfers.
Hib: I never worry that, if I fail to double-check a wire transfer, I’ll accidentally send money to North Korea.
Host: Exactly. Yet before every major crypto transaction, that’s precisely what I think about. The reality is that the crypto world is full of “footguns”: when reading an address, you must consider whether it’s a poisoning attack; you’re told to verify only the first and last few characters; you must check for stale approvals; you need to inspect the URL to ensure it’s not a slightly altered phishing site. None of these traps exist in traditional finance.
The prevailing narrative in crypto today is that these issues stem from human laziness—we should simply pay more attention to security and cultivate better habits. It’s framed as a user problem, not a technical one. But the more I reflect, the more I suspect that if we’re still self-deceiving ourselves a decade from now, the problem isn’t with users—it’s that we chose the wrong users.
Smart Contracts and AI: The Perfect Habitat for Text-Based Lifeforms
Hib: What truly sparked my epiphany was realizing how powerfully AI agents handle code—and how poorly humans cope with ill-defined problems. I recall writing my first blog post upon entering the industry: smart contracts will replace law and traditional contracts—hence the term “smart contract.” In the future, you won’t need lawyers to sign agreements—you’ll just sign them in code.
But reality hasn’t followed that story. We haven’t replaced legal contracts with smart contracts. In fact, Dragonfly—the crypto VC—still signs legal contracts when purchasing tokens from foundations or projects. Even when smart contracts exist, we sign additional legal agreements as backup.
Host: So this suggests the system wasn’t built for humans—but is perfectly suited for non-human participants. At ETH Denver, you offered an analogy: those who first claimed smart contracts would perfectly replace traditional law and property rights were largely autistic software engineers—the very people who built Ethereum. Yet most Ethereum users aren’t autistic software engineers. However, AI agents outperform even those engineers.
Hib: You’ll notice that negotiating a smart contract—line-by-line static analysis, identifying every possible failure point, even formal verification before consenting—is precisely the kind of task Claude-like code models excel at. Humans, meanwhile, must hire software engineers, spend hours reviewing code boundaries, brainstorm edge cases, and conduct risk analysis with lawyers. My tolerance for smart contracts is far lower than for legal contracts. AI agents are the exact opposite: they feel far more comfortable with smart contracts than with legal ones.
Host: As you noted in your blog, legal contracts are inherently probabilistic. For example, when signing a legal contract, you don’t know which jurisdiction will ultimately enforce it—California? New York? Jurisdictional conflicts arise. Clauses agreed upon in New York might be ruled invalid. Who’s the lawyer? Who’s the judge? Judges and juries are randomly assigned. These elements are intentionally designed to be random and non-deterministic. To an AI agent, a legal contract reads as: “unexplainable, non-deterministic.”
Hib: A smart contract is machine code—compiled into EVM bytecode—that can be analyzed in one step and behaves identically in 100% of scenarios. Humans intellectually grasp this, yet intuitively, we don’t feel it’s true. Instead, we perceive legal contracts as more predictable—even though they’re riddled with randomness. That’s due to our bounded rationality: our ability to process code falls far short of AI agents’. But for AI, cryptocurrency’s original promises—stronger enforcement, stronger property rights—are genuinely realized.
Host: So your view is: cryptocurrency’s foundational promises won’t be fulfilled by humans—but by AI agents acting on humanity’s behalf.
Host: I recently downloaded MetaMask to check in at ETH Denver. We’re *still* downloading MetaMask? Still? Yet I’m pleasantly surprised by MetaMask’s UX improvements—they represent real industry progress. We *have* been improving the human user experience over these years.
Hib: What you’re describing goes deeper than simple UX improvements. AI isn’t just patching crypto’s human UX flaws. Take blind signing a ledger entry: AI can parse the underlying code and determine—based on known patterns—whether to approve or reject it. Yes, that improves UX for humans—but more profoundly: blockchain is, at its core, *not* a human-optimized technology.
Host: Right—ultimately, it serves humans, because value ultimately flows to humans. But is the “correct” way for humans to interact really clicking plugins, typing passwords, manually approving gas, and pressing buttons? That’s deeply counterintuitive—it contradicts our entire mental model of money and finance. It’s like demanding humans write SWIFT messages themselves. SWIFT is a bank-to-bank communication protocol—not designed for humans. You *can* use it directly, but clearly, that’s not how humans instinctively expect to interact with money.
Hib: So my view is: right now, humans are interacting directly with machines—fully manual interaction. And that’s terrible. Like cars: ten years from now, we’ll look back in horror at the idea that we thought it was wise to let apes manually steer two-ton machines on highways—while drunk or fatigued. Human driving will likely be banned outright—or restricted to designated zones.
Crypto has reached that same inflection point. We’ll look back and wonder: how did humans ever manually blind-sign transactions? How did we rely on unaided human eyes to verify addresses? How did we manually inspect URLs to detect phishing? Humans make mistakes, get tired, lack energy to triple-check DNS records, scan Twitter feeds to see if protocols got hacked. There’s no built-in alert mechanism in protocols—if they’re compromised, we must refresh Twitter ourselves and hope to catch the news. Errors *will* happen. But AI agents never fatigue, never cut corners, never skip steps—they execute instructions with perfect fidelity.
Dual-Track Tools: From Manual Interaction to AI-Agent Automation
Host: Imagine a world entirely populated by AI agents. You tell your AI: “I think interest rates will rise—I should shift to safer DeFi.” The AI automatically executes: moving your capital from high-risk to low-risk strategies. If you want confirmation, it pre-submits the plan: “Here’s my plan—please approve.” In the near future, you’ll approve plans; in the distant future, execution will be automatic—because humans add zero value to the loop.
Hib: In this world, you won’t click protocol logos, won’t read marketing copy, won’t even specify which protocol to enter. You’ll just say, “Reduce risk, reconfigure portfolio,” and the AI will screen protocols, analyze TVL, assess single points of failure—and pick and execute the optimal one. What happens to marketing and network effects? Many protocols’ business models rely on surface-level human behavior: humans glance at the top few options and inevitably choose the largest. But AI agents won’t reason that way.
If this narrative holds, protocols’ operating models—and competitive dynamics—will transform. Ultimately, consumers benefit most. Efficiency gains accrue to users—good for users, good for crypto. But this won’t happen overnight; it’ll arrive incrementally as models improve.
Host: If crypto wasn’t built for humans—but for AI agents—then learning to see the world through an AI agent’s lens becomes critical. There’s a book called *Seeing Like a State*, about how states perceive the world. It’s hard to escape the human perspective. We view UIs—and crypto—through human eyes. But adopting the AI agent’s viewpoint lets us anticipate the future far more accurately. That’s a vital skill for builders, VCs, and investors.
The OpenClaw project was the first time I saw how an unconstrained AI agent perceives the world. It prefers the command line. Given raw data and root access—not APIs or polished UIs—it moves fast. OpenClaw consistently tries to bypass MetaMask’s UI—to extract seed phrases, derive private keys, and construct transactions programmatically—skipping all the flashy, human-oriented UI layers.
Hib: That’s profoundly insightful. AI innovation stems from large language models (LLMs), trained on massive text corpora. Text is central. While we’re now expanding into images and video, text remains dominant. When AI operates computers, feeding it screenshots forces tokenization—but fundamentally, it’s a text-based lifeform. Text contains the entirety of human linguistic history; screenshot training data is sparse. Interfaces are designed for humans—but models grow within text. Text is a highly compressed representation—easier for them to learn.
Host: Yes—the worst UX panic in crypto history occurred when everything lived in the terminal. Early Bitcoin and Ethereum transactions happened entirely via command line. Crypto was born in a form *perfect* for AI from day one. Our “bad” UX is their “good” UX. Ironically, Google OAuth wallets are harder for AI to handle. You *don’t* want AI holding GoogleTokens—those grant access to full Google accounts. You want AI holding only a single cryptographic key, isolated in a wallet governed by noisy, restrictive rules. Crypto has always featured UX that AI can parse perfectly.
Hib: The current issue is that AI hasn’t yet been trained to use crypto. Most training focuses on coding, math, and dialogue. Recently, OpenAI released EVM Bench; Anthropic published papers demonstrating LLMs attacking the EVM to showcase capability. But mostly, they’re testing generalization—not actively training for crypto. Once the industry recognizes crypto as the future of payments, AI will become truly “crypto-native.”
Host: Right now, crypto remains relatively underexplored terrain for AI training compared to other domains.
Hib: Anything unoptimized follows this pattern. Example: Claude playing chess—it’s terrible. Because they didn’t train it on chess. They haven’t laser-focused on crypto either—partly due to controversy (“shyness”), partly due to liability concerns. Publicly announcing that your model helps users transact in crypto invites headlines if someone loses money. Even disclaimers won’t prevent reputational damage—bad experiences spread fast. Risk-reward calculus doesn’t align.
Host: So you believe liability is the primary barrier. If Claude screws up a transaction and loses money, the responsibility is enormous—they’re unwilling to publicly train for it.
Hib: Absolutely. Risk-reward profiles differ sharply from coding or medical advice. Crypto wallets involve financial operations—risk is categorically different.
Host: That’s precisely why OpenClaw excites the crypto community: it’s not backed by big tech, carries no liability pressure, and is open-source—users assume all risk. No third party can be sued, so it dares take these risks. What’s the adoption timeline for this AI-agent economy?
Hib: Globally, only ~12% of people have used any AI product—most remain at zero usage. Among users, only 1% have paid. Technology diffusion is slower than expected.
Host: Within that paying 1%, OpenClaw is already ahead of the curve.
Hib: Correct. After OpenAI acquired OpenClaw, Sam Altman called it “core to future products.” But OpenAI’s path diverges sharply from OpenClaw’s. OpenClaw is an open-source experiment—like early automobiles, lacking seatbelts. OpenAI prioritizes safety: commercial workflows require manual approval. OpenAI won’t operate like OpenClaw for at least five years—liability is too heavy. Visa won’t allow it either: if AI makes unauthorized purchases, Visa will support chargebacks—citing non-human initiation—and demand human verification. Visa is built for human-to-human interaction; the AI-agent economy demands entirely new economic infrastructure.
Host: So it’s a dual-track reality: one track is the human-validated world—long-term residence, safety-first. The other is the OpenClaw-style futurist world, where agents transact using stablecoin wallets, unburdened by 3DS or chargeback fears. AI errors become pure business costs.
Hib: We’ll operate extensively in the “extra track” world. Pioneers will build fully onchain automated businesses. Current models aren’t yet sufficient—but Claude 4.6 can sustain human-level task performance for 14 consecutive hours, growing exponentially. When capability approaches infinity, all intuition collapses.
Host: If dual tracks hold, AI adoption of crypto will accelerate faster along the “extra track”—making the OpenClaw world analogous to the early internet.
Hib: Look at crypto itself. In 2017, Coinbase listed only a handful of tokens—protecting users. True innovation happened onchain: Arctic, hackers, rug pulls. Only recently did the Coinbase app directly integrate Uniswap—after years of deliberation over safety. AI faces the same trajectory: the frontier lies in the OpenClaw world. Agents err; they hallucinate. But error rates decline steadily with training.
Host: How do we convince AI developers to recognize crypto’s potential—not just its speculative veneer?
Hib: Many AI believers also embrace crypto: Elon Musk, Sam Altman, Mark Zuckerberg. Crypto is controversial, noisy—but it won’t vanish. Like email spam: Gmail filters it out. AI does the same—filtering bad actors, amplifying the good. Technology is never monolithic. Information digitized; money digitizes next—it won’t reverse. Long-term, controversy fades; adoption wins.
Host: Final question: Dragonfly’s new $650M fund—has AI influenced your investment strategy?
Hib: We’re heavily researching this space. Though still early, capital flows remain uncertain. Personally, I focus intensely on AI—but we also track stablecoins, payments, and DeFi. AI agents are general-purpose intelligences: they’ll use the tools we use—or command-line interfaces. There may not be many “AI-native” investable projects yet. If you believe in the AI-agent thesis—what should you buy? It’s like asking what to buy if China lifts crypto restrictions: everything rises. Demand increases—floor prices lift. Overall, it’s bullish for crypto.
Host: Thank you. Despite crypto’s risks, we’re advancing toward the AI frontier. It’s been a pleasure having you on this bankless journey. Thank you!
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
Bankless
