In the past, when discussing AI agents on the blockchain, everyone always thought "fully automatic = more advanced." Now I care more about where human oversight is absolutely necessary. For example, in the authorization/signature step, no matter how smart the agent is, it can easily be biased by prompt words, giving you an approval with unlimited limits... To put it plainly, if something goes wrong, it happens in an instant. Also, with cross-chain transactions, routing, slippage, and so on, the on-chain state changes instantly, and the agent might run along the most optimal path, but the most optimal isn't always the safest—especially when encountering strange liquidity or phishing contracts.

Another thing is whether "it should be done"—the agent can calculate this, but it can't understand your true hesitation or risk appetite. Recently, someone tried to link ETF capital flows, U.S. stock market risk appetite, and crypto price swings all together for analysis. It’s lively, but I still prefer to let the agent handle execution and leave the decision-making to humans, so at least I take responsibility if something goes wrong. For now, I only feel comfortable loosening the reins when the mechanism is clear.