Recently, everyone has been talking about AI Agents going fully automated on-chain, but I actually feel a bit apprehensive. Matching paths, slippage, gas—these it can calculate, but once it hits the mempool, encountering front-running, reordering, or someone else inserting into the same transaction—many times it's not about "enough computing power," but whether you're willing to withdraw, modify, or wait temporarily. I no longer believe in the idea that "full automation guarantees a win."

I now think at least three aspects still need human oversight: first, authorization/limits—don't let it recklessly grant unlimited wallet permissions; second, failure retry strategies—continuously bump gas or switch routes can easily lead to more losses; third, anomalies—price source fluctuations, RPC issues, cross-chain stalls on L2... Agents will just follow the script, but people need to watch out for "whether to stop."

By the way, on L2, people argue daily about TPS, fees, and who gets more subsidies. Frankly, for Agents, what matters more is transaction predictability: whether block production is steady, how often reordering happens, and how easy it is to trace failures and rollbacks. Anyway, I now prefer to go slower if it means I can clearly explain each step and feel more at ease.