Lately I’ve been seeing a bunch of AI agents claiming they can go on-chain and do work by themselves—automatically swap coins, auto-do all kinds of related actions, auto-vote, and so on. I’m really curious, but also a bit scared. Put plainly, in on-chain interactions, the part that most needs a human to step in is that “last confirmation”: who you’re authorizing, how big the allowance is, which route it takes, how much slippage you allow, and what you do if it ends up hitting a phishing contract… these issues really can’t be solved by a simple “I can reason.”

I thought the easiest way for an agent to go wrong would be in strategy judgment. But more often, it gets tripped up by the details: the same swap—if the path detours a little, if MEV takes a cut, the cost can end up looking completely different. And there’s also the problem of unlimited approvals—you may forget about it after a few days.

Also, lately the whole “attention mining” thing—social mining and fan tokens—is pretty noisy. My feeling is: attention really can be monetized, but in the end the bill is settled on-chain. One click to sign is basically exposing yourself to counterparties across the whole network… In any case, my current principle is: let the agent handle the dirty, tedious work, but wallet permissions and the exit button must be held in human hands.