3 Big Things That Happened in Crypto x AI This Week
Ethereum is quietly becoming the default home for the crypto and AI convergence. Three deep-dives with Dragonfly, Paradigm, and Nat Eliason this week made one thing clear: the agent thesis is less about creation and more about execution.
Here's the breakdown.
Agents Are Tools, Not Thinkers
The most consistent theme across all three conversations - agents are amplifiers, not idea generators. They pull from the center of their training data, which means genuinely novel business thinking is outside their reach. The edge still has to come from humans who've built real knowledge over time.
@nateliason's
@FelixCraftAI is a good example of how this works in practice. The agent handles execution while the underlying intelligence - years of productivity and knowledge management expertise - comes from Nat himself. The marketplace it runs sells exactly what agents can't produce on their own.
Same logic applies to trading. An agent can execute a proven strategy faster and more consistently than any human. But it won't find the alpha. That part is still on you.
New toolkits dropped this week to support this kind of specialized agent work -
@austingriffith's Ethereum dApp toolkit,
@OpenZeppelin's smart contract integration, and
@clanker_world's token deployment tools.
Agents and Crime Are a Natural Fit
@hosseeb made the point bluntly - agents are uniquely suited to operating in grey areas. No sleep, no jail time, persistent execution. Crypto is already the front line of this because it's permissionless by design.
The threat isn't abstract. As agents start interacting with more unknown counterparties, bad actors will impersonate tools, inject themselves into workflows, and probe for exploits continuously.
ERC-8004 is one of the more interesting responses to this - onchain identity and reputation scoring for agents, so you can verify who you're actually dealing with before any interaction happens. As agent activity scales up, this kind of infrastructure stops being optional.
AI Is Getting Genuinely Good at Security
@0xalpo's EVMBench paper with OpenAI put some real numbers on something the space has been speculating about for a while. Six months ago AI models caught around 12-13% of critical smart contract bugs. That number is now above 70% with the latest models. Top human auditors are expected to be outperformed within the next six to eight months.
The insight from
@hosseeb that reframes this - crypto's notoriously bad UX, all text and code, no visual noise, is actually ideal training data for AI. The environment that frustrated users for years is the same environment that makes agents exceptionally capable.
@pashov, who has audited Aave, Pump, Ethena, and Uniswap, released an open-source Solidity auditing agent last week. If the defender side deploys these tools faster than attackers do, AI becomes one of the most important security layers the industry has ever had.
The trust ceiling is the real bottleneck for crypto growth right now. Better AI auditing is one of the most direct ways to raise it.