My phone just popped up a red dot again: an agent "has automatically completed cross-chain + deposit and lending," which looks pretty convenient, but I still manually opened the on-chain record to verify... To put it simply, the biggest risk for agents now is boundary conditions: granting too much permission, routing through strange pools, executing even when gas prices spike, or the bridge getting stuck without a receipt. And there are those new L2/L1s that, once incentivized, pull TVL aggressively; it's no wonder old users complain about "mining, withdrawing, and selling." The agent will only follow the rules; if no one watches it, it might automatically take the last step. Anyway, my current fallback plan is three things: keep permissions as minimal as possible, verify each step (tx hash matches), and stop immediately if an anomaly occurs, preventing it from continuing on its own.