Recently, watching AI Agents go on-chain to interact, it’s indeed more convenient, but when it comes to critical steps, someone still needs to take responsibility. For example, during authorization, many contracts are written in a complicated way, and the Agent just signs if it sees “usable,” so it’s not surprising if assets are dragged away with unlimited limits later; another example is cross-chain/bridges—if something goes wrong, you need to judge whether to wait, roll back, or admit defeat, but machines will just keep hitting the wall according to the process. Also, during the period when new L1/L2s offer incentives to attract TVL, old users complained about “mining, selling,” but I actually think it’s more necessary to have someone watch: Agents will work diligently to harvest, but they won’t ask you whether this is a one-time profit. I set up a small position for it to run, and the rest I still confirm manually—slower but I can sleep peacefully.