Recently, I've seen everyone talking about AI Agents going on-chain to do work, and it feels pretty cool, but when it comes to actually using them, someone still needs to cover the backend. To put it simply, Agents will execute, but won't be responsible: the signature, who to authorize, how much limit to set—if any of these are clicked incorrectly, it's not something a "retract" can fix. Especially now, with phishing links being so common, pages made to look exactly like official websites, even the smartest Agent could be fed a fake address... I actually prefer to take it slow and go through the key steps myself.

Another aspect is "intention" and "boundaries." For example, if it helps you rebalance your portfolio or do DeFi, and suddenly there's a large slippage, the pool gets drained, or the contract is upgraded—these low-probability, high-risk pitfalls—human intuition (or suspicion) is quite useful: stop first, find another way, or just don't do it. Recently, hardware wallets are also out of stock, which further shows that cold signing isn't just for show; if you can use a hardware wallet, do so. Think of the Agent as a diligent assistant, not a driver. Anyway, my current principle is: let it run errands, but don't let it hold the keys.