These days, I’ve seen a few demos of AI agents “automatically working” on the chain again. It’s definitely convenient, but my biggest concern is still the lack of a safety net: once the steps of authorization, route switching, and signing go wrong, you won’t even know who to blame later. Especially since allowances are often unclear, and they might casually set an unlimited limit—basically, handing over the keys without watching the door.

My current way of judging whether an agent is reliable isn’t how fast it runs, but whether it pauses to ask during fee rate/slippage/route changes, or at least shows you the differences openly. For protocols that secretly change parameters, I really take screenshots and archive them… Otherwise, if one day I find “it’s more expensive again,” I’ll just have to suffer the loss myself.

Recently, with new L1/L2 incentives to boost TVL, in environments where everyone’s rushing to mine, transfer, and sell, agents are more like accelerators: they can rush in quickly and also exit quickly. But if they get stuck at a bridge, face low liquidity, or if the incentive rules are temporarily changed, you still need someone watching over it, with your hand on the revoke authorization and pause buttons. Anyway, I currently don’t dare to entrust them fully; at most, I let them assist.