Recently, everyone has been hyping up AI Agents that can run fully automated on the blockchain, and I was tempted too, but the more I look into it, the more I feel it still needs human oversight. For example, during authorization, the agent might casually give itself an unlimited limit to "complete the task," and frankly, if something goes wrong, you're the one paying the price; also, with cross-chain swaps or pool migrations, a change in the route can cause slippage, and that "gas saving" on the blockchain can sometimes turn into a trap. Then there are contract upgrades and governance proposals—just because the agent can read the words doesn’t mean it understands the risks, so in the end, you still have to click the confirmation yourself.

The past couple of days, the unlock calendar was brought up again and it was shocking. I definitely don’t dare let the agent chase hot trends recklessly anymore. I’d rather try small amounts as experiments, and patch issues if they arise: reduce authorization limits, whitelist addresses, stop if failures occur… small fixes and adjustments, at least I can sleep peacefully.