Recently, watching AI agents go on-chain to do work has been quite lively, but honestly, someone still needs to backstop it. The signing step always requires supervision: authorization scope, whether it's unlimited, whether the contract address is correct. Even the smartest agent can be fooled by phishing pages. When the network is congested and gas prices spike, it might just push through, ending up paying more and getting stuck for a long time. In such cases, manual intervention is needed to stop it or switch routes. The same applies to strategy layers—when encountering meme or celebrity shoutouts that shift attention, agents might only chase data hotness. Veteran players still remember the phrase "don't take the last baton"... Anyway, I currently have it doing monitoring and drafting orders, but the final confirmation still comes from me—better to play it safe for now.