These days, everyone is talking about AI Agents automatically running on-chain, but when it comes to actually executing, someone still needs to take responsibility. For example, choosing routes across L2s, the Agent can calculate fees and time, but if there's bridge congestion, the sequencer acts up, or gas suddenly spikes, it’s easy for it to "force through" according to plan, ending up costing more and getting stuck. And the most critical part is authorization—Agents like to approve infinitely with one click, but honestly, all the risks are left to you.

Recently, hardware wallets are out of stock again, and phishing links are everywhere, so I’m even less willing to hand over signing rights. My approach is: the Agent only monitors on-chain status and generates plans; the final step of confirming, verifying contract addresses/authorization scope, and adjusting limits must be done by a person. It’s convenient, but don’t skip thinking.