Recently, I've seen a bunch of people bragging about AI Agents placing orders on-chain and off-chain by themselves, finding routes on their own, managing positions independently... I admit it saves effort, but honestly, the most critical part that still needs human oversight is the "last mile." For example, during authorization, an agent might give you an unlimited approval to save on gas, or sign off on fake frontends/fake routing too. Or during congestion, issues like mempool stalls, being front-run, or re-bundling—agents will just retry based on parameters, but humans need to monitor and adjust strategies or simply cancel orders. And then there are the bridges and message delays across chains and Layer 2 solutions—don't think that now, with Layer 2's hype over TPS and fees, when your assets get stuck halfway through, the agent can only "wait for confirmation." I need to be reminded: automation should be as automatic as possible, but signing, approval limits, emergency stops, and rollback in case of anomalies are best left to humans to confirm with a click. Anyway, I’d rather be slower now than risk stepping into another pit.