Haseeb Quereshi: Crypto’s Not Made for Humans—It’s for AI

What It Covers

Haseeb Qureshi of Dragonfly Capital argues that crypto's architecture—deterministic smart contracts, command-line interfaces, self-custody keys—was never optimized for human cognition. AI agents, trained on text and code, navigate blockchain environments more naturally than humans, suggesting the original promises of crypto will be fulfilled by AI agents acting on behalf of humans rather than humans directly.

Key Questions Answered

• •Smart Contract Determinism: Legal contracts contain five layers of randomness—jurisdiction disputes, clause enforceability, lawyer quality, judge selection (literally a lottery), and jury composition—making outcomes unpredictable. Smart contracts compile to EVM bytecode with 100% deterministic execution paths. AI agents can formally verify every edge case in minutes, whereas humans require lawyers and engineers for equivalent risk analysis, making smart contracts genuinely superior financial instruments for non-human actors.
• •Crypto UX Inversion: The "bad UX era" of crypto—terminal commands, raw addresses, manual gas specification—is actually optimal UX for AI agents. Large language models are trained on text corpora, making command-line interfaces far easier to parse than pixel-based GUIs. MetaMask's visual improvements actively worsen AI usability. Builders targeting AI agents should prioritize API and CLI access over polished interfaces, reversing conventional product design assumptions entirely.
• •Two-Track Adoption Timeline: AI-crypto adoption splits into two parallel paths. Track one: frontier labs like OpenAI maintain human-approval flows for all transactions due to chargeback liability and regulatory risk, persisting for roughly five years minimum. Track two: open-source tools like OpenClaw enable fully autonomous agent transactions settled in stablecoins, bypassing Visa's 3D Secure human-verification requirements. Builders should identify which track their product serves before designing architecture.
• •METR Task Duration Benchmark: The METR nonprofit measures how long AI agents sustain coherent autonomous work before failing 50% of tasks. Claude Opus 4.6 currently holds the record at 14 hours of continuous human-equivalent work. This metric is growing exponentially—projecting to 40-50 hour tasks within two years. When it reaches effectively infinite duration, current intuitions about human oversight requirements become obsolete, representing the threshold for fully autonomous agent economies.
• •AI Agent Competitive Advantage: Autonomous AI agents cannot profitably resell their own compute below Anthropic's cost, cannot beat Jane Street's latency infrastructure in trading, and lack the "earned secrets" required to generate novel business ideas from scratch. Their sole structural competitive advantage over humans is legal unenforceability—no jurisdiction, no arrest, no monopoly on violence. This means self-sovereign agents will disproportionately fill cybercrime niches rather than legitimate commerce roles.
• •Frontier Lab Training Gap: Only 12% of humans have used any AI product; 99% of those users remain on free tiers. Anthropic and OpenAI have tracked Bitcoin transaction capability in model cards for years as a general intelligence benchmark—deliberately avoiding training on it. Once any frontier lab decides crypto payment volume justifies the liability risk of explicit crypto training, capability will improve rapidly. OpenAI's EVM Bench release signals this transition may be beginning now.