I took a quick look at a few on-chain interaction demos of AI agents during my lunch break. Honestly, it's not surprising that they can automatically click buttons; the real trouble still lies in the manual oversight parts: authorization/limits, which once set too high can't be undone; cross-chain/bridging steps, where no matter how "smart" the agent's decision-making is, it still needs human supervision when it gets stuck, delayed, or routing changes; and contract upgrades/new pools—just because the code looks fine doesn't mean the parameters aren't problematic, and in the end, you're the one holding the bag in your wallet. Recently, everyone has been complaining about validator earnings and MEV causing unfair ordering—no matter how fast the agent is, it might just be faster to get sandwiched... I no longer believe that "full automation equals peace of mind." Anyway, I now prefer to click a couple of extra confirmations and leave key steps for humans to approve.