Lately, there have been quite a few demos of AI agents interacting with on-chain activity, and they run pretty smoothly. But honestly, I still don’t dare to fully delegate everything. Once it gets “too confident” about the signature step, the authorization limits, and the cross-chain/bridge route choices, it’s easy for things to go wrong—ultimately, the person still ends up taking the blame. And on top of that, there’s always some MEV gossip in the mempool: things like slippage, reordering, and sandwiching it can calculate. But when it runs into abnormal liquidity or little traps in the contract, you still need a human to step in and pause.

Why am I so calm? One small habit: for any new agent or any new contract, I first run it with a small account, a small amount, and a limited-time authorization. Then the next day, I check the on-chain records and the authorization list. Only if everything looks fine do I slowly loosen the restrictions. Recently, there’s been another wave of meme coins and celebrity call-outs. Newcomers are very easily pushed along by attention—I just treat it as noise… first, lock down permissions, and don’t take the last step.