Lately, I’ve been seeing everyone talk about AI Agents interacting with on-chain activity. Some people are rushing to hype “fully automated,” while others have started nitpicking the security details. After trying it myself, there are actually quite a few parts where a human has to stand in for you. The authorization step is the one I fear most—one slip of the hand is all it takes. And in the signature pop-up, you still need to be absolutely clear on what permissions you’re granting. No matter how smart the Agent is, it can’t just “take the blame” for you.

And for workflows like cross-chain/exchange, once the path gets long, it’s easy to get stuck. It’s just a simple confirmation click, yet you end up waiting in a queue for ages. In the end, I can only refresh/retry—basically running the whole mindset retry loop along with it…

Anyway, my current rules are pretty old-school: let it run for small amounts, but for large amounts I have to manually go through the permissions and the address first. I’d rather move slower than have one mistake get blown up into a whole screen of red pixels.