Recently, everyone has been hyping AI Agents as if they can “go to work on the chain” for you… And sure, it really can help you click buttons, handle bulk interactions, and basically farm those testnet tasks—but when it comes to the crucial step, you still need a human to backstop it. For example, the signature step: in the pop-up, what exactly is it authorizing, is it unlimited authorization, is the address correct—no matter how smart the Agent is, it might still just click “to complete the task” “for the sake of finishing,” and I don’t want to bet real money on its confidence.

As for the testnet incentives and points system, everything right now is being propped up by expectations. People in the group ask every day whether the mainnet will issue tokens—I get tempted too—but the more tempted I feel, the more I need someone human to stay cold and calm: does this project actually have real use, is the interaction just pure grifting, and what if the rules change? Anyway, my current backstop strategy has only two parts: permissions are given at the bare minimum, and I take a screenshot of every interaction for the record (yes, I’m once again collecting those weird “high-headed evidence” screenshots, the same vibe as bizarre K-line patterns), so that if I look back in a couple of days, I don’t end up with only one line: “Back then, I also didn’t know how I ended up clicking it.”