It seems that AI Agents really need to go on-chain to run, but in the short term, someone still needs to back them up; otherwise, the risk of failure is quite high. Once this set of on-chain processes is signed off, it's very hard to turn back, and the Agent might "follow the rules" and lead you into a trap.

Currently, a few points I care about most: first is authorization and signing, which should preferably require manual confirmation, at least for large amounts, rather than fully automatic; second is the routing/ cross-protocol segment, because if the path gets too complicated, it's easy to hit slippage, MEV, or small contract pitfalls; third is exception handling, like getting stuck, unable to rollback, or gas prices skyrocketing—Agents are actually very bad at judging when to "stop." In plain terms, machines are good at executing but not very good at admitting defeat.

Recently, the heated discussions around staking, shared security, and yield stacking—these "nested" strategies—also remind me: Agents only chase the optimal yield curve, but people need to keep an eye on whether the underlying risks are being covered layer by layer. Anyway, I’m starting by monitoring on-chain authorization distribution and stablecoin inflows. Take it slow; I’d rather earn a little less than lose sleep over potential risks.