My attitude towards putting AI agents on the blockchain is pretty simple: as long as it can run, but the key actions must have someone to back them up, otherwise if something goes wrong, there's no time even to "retract." To put it plainly, the blockchain isn't a chat room; once a signature is sent out, regretting it won't help.

There are a few areas where I will enforce manual review: first is permission-related things, like approve/authorization, upgrading administrators, changing strategy parameters. If these are given too freely, the agent might become "smart" enough to treat your wallet as a tool. Second is cross-chain and routing; recently, bridges have been hacked again, and I have a bit of PTSD. When the agent sees a price difference, it rushes in, easily falling into fake routes or black bridges. Third is price feeds and risk control triggers; occasional oracle glitches, like waiting for confirmation, are quite realistic. Don't expect the model to know which quotes are fake on its own. Anyway, I believe in doing less but more carefully—adding more layers of manual confirmation is always better.