Lately I've been seeing a bunch of AI Agents claiming they can "fully automate on-chain surfing," and I feel both excited and amused... In reality, for the critical steps, humans still need to be the safety net. For example, during authorization, the agent might only see "can execute," but humans need to see "who was given access, how much, and whether it can be revoked." Otherwise, an unlimited authorization could turn you into a public water faucet. Another example is governance voting—agents can write decent proposal summaries, but they might not pick up on conflicts of interest or subtle voting strategies. In the end, I still have to keep an eye on the comment section for arguments and disputes.

And regarding deposits and withdrawals, recently some places have tightened or loosened taxes and compliance, making everyone's expectations swing like a roller coaster. Agents will just follow the rules, but I need to be reminded: if the rules change, pause first. Don’t push hard during the "seems possible to transfer" window. My current principle is: agents handle the errands, I handle the brakes. Who will remind me daily not to get itchy and grant reckless permissions... I need to be reminded.