I tried once to have an AI agent help me run a set of on-chain interactions, and I found that the only part that was really convenient was the "click click click" part: checking balances, looking at gas, filling in parameters according to the process—those were pretty stable. But when it comes to decision-making steps, you still need a human to cover the bottom line, like how much authorization limit to give, what tokens are included in the routing, whether the contract has been replaced with a different implementation, and those pages that say "profits are attractive" but the underlying assets are wrapped three times... To put it simply, it will follow instructions to complete tasks but won't take responsibility for whether you trust it or not.

Recently, everyone has been comparing RWA and US bond yields to on-chain yield products. I also look at them, but in the end, I still come back to one sentence: the extra yield on-chain usually corresponds to unseen risks. Anyway, I now just let the agent handle execution, and I only focus on two things: where the money goes and who gets the permissions. It’s okay if it’s slow—just make each interaction clean.