Recently, I’ve been looking at things where AI agents interact with the blockchain—and they look really beautiful: automatically finding pools, automatically rebalancing, automatically voting… But let’s be honest: I still feel there are a few steps that must be backed up by humans. For example, when it comes to authorization, no matter how smart the agent is, it can’t stop you from granting an infinite allowance to a strange contract with your own hand. And for example, with the question of “whether to chase yield”—it can calculate, but it isn’t afraid of drawdowns; you are. Later, I found that the incentive design is actually the easiest thing to mess up: the agent can be fed like a brush-upter with subsidies, very similar to the inflation + studio + token-price spiral you see in chain games. Once the logic goes wrong, automation only speeds up the collapse. In any case, my current approach is: test with small amounts, tighten permissions, and confirm key actions yourself—slower, but at least I can sleep at night.