These days, I've been looking at the automated order placement on the AI Agent chain again. It’s really convenient, but honestly, no matter how smart it is, it can only follow scripts: once you give too much authorization/limit/routing, encountering phishing contracts or bizarre tokens, problems can happen in an instant. I still manually review permissions and whitelists myself, at most letting it handle small, reversible actions.

The turning point is that many pitfalls on-chain are not something "computing power" can avoid, such as temporary maintenance of cross-chain bridges, RPC outages, MEV front-running. The world the Agent sees might be even slower than yours, and in the end, it still takes humans to pause, change parameters, or cut losses.
By the way, over on L2, they’re arguing about TPS/fees/subsidies again… It’s lively, but I’m actually more concerned about whether these fundamental infrastructures, once shaken, will cause the Agent to treat me as “normal executable” and keep rushing forward. Anyway, my current principle is: automation can accelerate things, but the brakes must be in my hands.