Recently, I’ve been watching a bunch of demos of AI Agents that automatically carry out on-chain operations. It’s definitely convenient, but the truth is that when it comes to actually paying out money, you still need a human to cover you. For example, during cross-chain operations—when there are bridge delays, stuck blocks, or routing temporarily changes—the Agent will just keep hammering away according to the script. In the end, if the settlement is off by a little, you can only own that outcome yourself. And with things like approvals, swapping pools, slippage, and so on, once the on-chain environment changes, it’s easy for them to “follow the instructions but get it wrong.”

Right now, at most I let it run monitoring and simulation. The order placement and the exit/withdrawal buttons are still something I click myself, especially when major mainstream public chains are upgrading or undergoing downtime maintenance—back in those days, everyone in the group was speculating whether projects would migrate. I’m just going to check first whether the settlement can be smooth.

If it’s complicated, I treat it as the enemy—fewer steps is better.