Recently experimenting with AI Agents running on-chain interactions, it really saves effort, but honestly, someone still needs to oversee it. For example, with authorization, agents can easily give you unlimited permissions just to "complete the task," so now I make it a point to always check the spender and limits at the end; also, for signatures/transaction simulations, no matter how convincing the agent's reasoning is, I won't feel secure unless I run a simulation myself and verify what contracts are actually being called. Gas fees and slippage are the same—agents will "smartly" help you save money, but during congestion or routing issues, they can also get you stuck halfway.

Additionally, when there's talk in the community about stablecoin regulation, reserve audits, or rumors of de-pegging, and emotions run high, the agent will also "advise immediate action," which is very dangerous... My current rule is: any operations involving main holdings or stablecoins swapping back and forth must be manually confirmed, better to be slow. Automation can run tasks, but risk should not be automated. Talk again next time.