Lately, I've been watching everyone talk about AI agents running on-chain strategies with a single click, and I still feel a bit anxious… They really can do the work, but I think at least three areas still need human oversight: First, during authorization/signing, if something goes wrong, you won't even know who to blame; second, the steps for cross-chain/bridges, where the agent might only look at the "success rate," but humans need to consider the "worst-case scenario"; third, when collateral positions face extreme volatility and correlations go haywire, someone needs to decide on stop-loss or leverage reduction, otherwise it just automatically gets sent to liquidation zones.

Additionally, Layer 2 is now competing over TPS, fees, and subsidies, and the agent might just naively pick the "cheapest" option, resulting in insufficient liquidity depth and slippage eating up the small savings on fees… Anyway, I’d rather do some manual confirmations, slower but more reliable.