Lately, I've been watching AI Agents go on-chain to do tasks, and it feels like they run pretty fast, but if you want to scale up funds, someone still needs to backstop. For example, during authorization, the machine checks "can it execute," but humans need to watch out for "what if it's a phishing attack or the permissions are too broad"; then, for cross-chain or pool swapping, if slippage or routing changes, the Agent might follow the script, but humans will pause accordingly. And the most annoying part is when contracts are upgraded or parameters are changed, it may not realize that "the environment has changed."

These days, I also see people comparing RWA and U.S. Treasury yields to on-chain yield products. Basically, on the surface, it's about returns, but underneath, the sources of risk are different. Agents are even less good at judging "what exactly is supporting this yield."

Yesterday, I even considered uninstalling a few automation tools... to save trouble, but I was also worried about missing opportunities. In the end, I kept them, with limits + position splitting + periodic authorization revocation, watching slowly, no need to rush.