Recently, I've seen everyone hype up AI Agents that can help you do on-chain work with one click.

I'm also tempted, but honestly, there still needs to be a "human fallback" option.
For example, during the authorization step, the agent might think it's convenient and give you a high limit.
Now I prefer to go slower, confirming each step one by one.
And for cross-chain transfers, after the recent bridge theft incidents, I really don't dare let it automatically jump from Chain A to Chain B.
At most, it can help me calculate the route, but I still do the final click myself.
Then there's the occasional glitch with oracle price feeds.
The group is all waiting for "confirmation," and I wait along with them.
No matter how smart the agent is, it can't take the blame for you.
Once I encountered a task where I didn't understand what strange message I was supposed to sign, and the page looked complicated...
I just closed it, and the next day I found out it was a phishing site.
Luckily, I chickened out at the time.
Anyway, if it can be automated, let it be automated.
For things involving authorization, cross-chain transfers, or signatures, doing it manually at a slower pace is more stable and safer.