Recently, as AI Agents on-chain interactions become more and more smooth, I actually care more about "which steps must have someone to take responsibility." The most typical case in authorization: no matter how smart it is, it might lead you to sign an approve you don't understand, and retracting permissions afterward is a hassle for everyone... Also, cross-chain/bridges, with more routes, are more likely to be "optimized" routing to strange places, making it hard to trace issues. Then there are irreversible choices like contract upgrades and governance votes; Agents can give suggestions, but ultimately, the person pressing the button should probably do it themselves. By the way, I want to complain that many on-chain data tools and tagging systems are said to be lagging or misleading, which isn't without reason. If the "address profile" fed into the Agent is biased, everything afterward will go off course. Anyway, my current approach is: let it check, let it compare, but I review the signatures and permissions myself before proceeding.