Recently, I've been seeing everyone hype up AI Agents working directly on the chain. Honestly, I'm quite tempted too, but when it comes to actual implementation, there are a few steps that require human oversight. For example, whether it needs to be authorized first, how broad the authorization scope should be, whether it can automatically renew, and so on. Once you sign the wrong agreement, it's not as simple as "retract and restart"; it's more complicated. There's also the routing/cross-chain part—when the Agent sees "cheap" and "fast," it rushes in. But it might not consider whether the bridge contract or target pool has recently been patched or if there have been any abnormal events. It might not stop to think twice. Now, I ignore Layer 2's arguments about TPS/fees/subsidies as background noise. In actual experience, the easiest way to go wrong is still with permissions and contract updates. My approach is simple: the Agent is responsible for execution, I handle the final confirmation and check security announcements, and I don't let it expand the position on its own. That's it for now.