Recently, I’ve been hearing everyone hype AI Agents to automatically handle tasks on-chain. To be honest, I’m pretty tempted—it saves effort. But once you actually run it, there are still a few parts that must be covered by humans: First is the boundary condition of “whether I should do this transaction or not.” The Agent is good at calculating routes and fee structures, but it doesn’t take the blame—or the risk—if “this project/this pool suddenly pulls liquidity.” Second is authorization and signing: if the permissions are too broad, I won’t be able to rest easy; if they’re too narrow, it just keeps failing. In the end, a person still has to watch and properly lock down things like the whitelist, limits, and expiration times. Third is exception handling—getting stuck in pending, routes suddenly getting more expensive, MEV getting in the way… it might still be “optimizing,” but only a manual stop from you can stop the bleeding.

Memes and celebrity-style attention cycles are even more troublesome. What the Agent sees is that on-chain trades are hot; what you have to consider is whether “I’m the last one in the line.” I need to be reminded: automation can only help me do fewer errands—it can’t save me from making fewer stupid mistakes. First, clearly write down the maximum amount I’m willing to lose, and then we’ll talk.