Recently, everyone’s been hyping that AI Agents can automate on-chain operations—it really does save time. But when it’s actually time to run it, I still think someone needs to be there to cover the risk. For example, in the permissions/limits area: if you set them too high, one careless slip and it turns into “automatically delivering warmth”; if you set them too low, it easily gets stuck. And as for signatures—let’s be honest, machines only confirm according to the rules. Humans first need to write the rules correctly, and then keep an eye on anomalies; otherwise, when it hits a phishing contract or fake routing, the Agent will rush in just the same.

Also, nowadays, big on-chain transfers and any movement of an exchange’s hot/cold wallets are immediately interpreted as “smart money is coming in.” If the Agent also treats that kind of signal as input, it’s pretty easy to get played along. Anyway, I’d rather have it act as an execution tool, while humans handle judgment and clean up the mess—at least if something goes wrong, you know who to blame.