Recently, I saw AI agents going on-chain to do work, which is indeed convenient, but honestly, someone still needs to cover a few steps: authorization is the most critical part, the agent will just think "approval is needed to complete the task," I just want to say "please don't give me unlimited authorization"... Now I see unlimited, I manually change it to a small amount, or even first approve 0.01 to test the waters. There are also slippage and routing issues with cross-chain/swapping pools, the agent chooses quite smartly, but when liquidity is thinner, it’s easy to feed itself to MEV, and in the end, someone still needs to click confirm, or simply set a limit price. On the macro side, they’re talking about rate cut expectations, the dollar index, risk assets sometimes rise and fall together, sometimes move inversely, the agent has no emotion management, and won’t scold itself when losing, I would. Anyway, letting it handle chores is fine, but for signing and the question of "can it afford to lose," I’ll still take care of those.