Recently, watching AI agents interact on the blockchain, many people are excitedly saying "It will be fully automated in the future."

I, on the other hand, feel a bit nervous: I see simplicity as a trap.
Authorization, limits, contract addresses, upgrade proxies—these steps really can't be fully delegated to the model for a one-sentence decision.
If something goes wrong, who takes responsibility?
To put it plainly, an agent is most like a diligent assistant, but the "final confirmation" must be done by a human, especially when it comes to signing and changing parameters.

And then there's the impulse to let the agent find the optimal path on its own—chasing high yields and spamming tasks, which is quite similar to the inflation + studio + coin price spiral in blockchain games.
Once the incentives become distorted, it can quickly collapse.
Anyway, my current approach is: read the original text when possible, minimize authorizations when possible, let the agent handle errands, but humans must guard the gate.