Lately, I’ve been watching AI Agents automatically run on-chain—it’s definitely convenient, but honestly, someone still has to be there holding the lid. I never dare to fully let go of the signing step: no matter how smart it is, it can’t tell the difference between a “real official website” and a “phishing site that looks like the official one.” Especially these days when phishing links are flying everywhere and hardware wallets are out of stock—if something really goes wrong, even switching to a more secure device can be a hassle.

And then there’s the authorization/limit part: the Agent often takes shortcuts and just sets unlimited allowances. I usually manually reduce them and then, afterward, revoke them right away—otherwise it’s like pouring the whole salt jar into the pot: once it’s done, trying to save it later is hard.

Finally, there’s the “when it’s ready to come off the stove” timing—automatic take-profit and stop-loss sound great, but when the chain gets congested and slippage suddenly gets bigger, you still need someone to watch it so it doesn’t get overcooked and burned.

Forget all the extra tricks for now. Let automation do what it can, and keep a little human “hands-on” backup for the key steps—it just feels more solid.