The more I watch AI Agents, the more they feel like “hands” rather than “brains.” They can help me monitor the market, compare prices, place orders in batches, and knock out a bunch of on-chain steps—but when it comes to the part where you need someone to step in as a fallback, it still has to be a human: don’t just hand out/approve permissions and signatures casually (one approval can keep you up all night), irreversible actions like cross-chain/bridges, new contract pools, and weird routing. No matter how smart the Agent is, it can turn “saving on gas” into “saving by cutting out your principal.”

Lately, everyone has been benchmarking on-chain yield products against RWA and US Treasury yield. I’ll have the Agent pull and align the data too, but that final question—“what exactly is propping up this yield?”—still has to be checked by me: is it interest, incentives, or just emotion?

Anyway, my current rule is: let it handle the repetitive work, and keep the buttons that carry high regrets for myself.