Recently, everyone has been talking about AI Agents going on-chain to do work, describing it like fully automated managed life...

I'm actually more concerned about which parts still require human oversight.
For example, the authorization step—once the signature is too broad, no matter how it "optimizes" for you later, there's no way to stop it;
And before contract interaction, confirming the address/chain—no matter how smart the Agent is, it can still be fed incorrect data, and one wrong input is enough to cause a problem.
Then there's strategy switching and stop-loss mechanisms—during extreme market conditions, it might just execute mechanically, so you need to have a "pause" button yourself.
Right now, the narrative is quite hyped, but those truly focused on security are mainly watching out for permissions, whitelists, and rollback plans—boring details, really.
In short, an Agent can save time but can't take the blame if something goes wrong.