I just turned off the on-chain script for a small Agent again. Running it is pretty smooth, but when it comes to signing, it still requires human oversight. Otherwise, it sees "cheap gas" and starts acting recklessly, completely ignoring small but deadly details like slippage and approval scope. To put it plainly, machines only execute based on conditions; they won't worry for you.

What I can now delegate to it: checking balances, calculating positions, scheduled rebalancing, and turning profit/risk alerts into notifications; but involving unlimited approvals, cross-chain bridges, contract upgrade addresses, and emergency withdrawals (especially sudden on-chain freezes or rollbacks) still require human oversight. Recently, before and after major public chain upgrades, everyone speculates about migrating ecosystems. My approach is more old-fashioned: I pre-write the withdrawal paths and alternative pools, and the Agent only handles notifications and simulations. The person who actually presses the confirm button must be me.