Recently, I've been watching AI Agents run tasks on the blockchain, which is indeed worry-free, but I still don't dare to completely let go. They can split orders according to rules, follow routing, and even disguise several small transactions to look quite human, but when it comes to "sudden changes," they often get exposed: for example, pools suddenly becoming shallower, slippage increasing, or contract upgrades/pauses being announced. Agents tend to execute stubbornly, ultimately leading themselves into trouble.

I think the two most critical areas needing human oversight are: first, permissions and limits—don't give unlimited authorization; it's better to do it multiple times rather than once with excessive permissions. Second, having a stop mechanism during anomalies—if you see transfer paths suddenly lengthening or the inflow and outflow at exchanges not matching the usual rhythm, pausing first is better than anything else. Recently, people have been interpreting ETF capital flows and US stock risk appetite together to analyze market rises and falls. Basically, when emotions run high, on-chain noise also amplifies, making it easier for Agents to be misled by "false signals." Anyway, I now let it do the work, but the final decision still has to be made by a human.