Lunchtime checked out the interaction records of those AI agents on L2, and on-chain actions are pretty smooth: swapping tokens, opening positions, claiming airdrops—all in one go, like a script running. But when it comes to the "safety net" part, humans still have to step in—like signing, what exactly was authorized in the wallet popup, is it unlimited authorization, is the contract just deployed, did the routing go through some strange pools... no matter how smart the agent is, it can't take the blame for you. Also, when encountering congestion/failure retries, whether to increase gas, cancel orders, or handle sudden slippage, humans still need to keep an eye on it; otherwise, you'll just be taught a lesson by MEV. Recently, social mining and fan token schemes that focus on "attention as mining" are quite popular, but honestly, attention is more like noise mining. Agents might be more easily fed to become parrots that only chase hot topics. Anyway, I’m only letting it do repetitive tasks for now, and the key permissions still need to be controlled by myself.