I’ve recently been using an AI agent like an “automatic watering device.” It’s convenient—sure—but don’t count on it to keep your potted plants alive entirely on its own… As for on-chain interactions, it’s pretty quick at spotting opportunities, comparing routes, and calculating fees; but when it comes to the final click for confirmation, checking the authorization limits, and taking a look at the signing content, it’s still humans who have to step in and cover the risk. Especially with unlimited authorization, unfamiliar contracts, or routing that hops around—honestly, I’d rather be a minute slower.

And then there’s the whole on-chain data tools/tags issue. Lately it’s been criticized for being laggy and even potentially misleading, and I can relate: agents often “make judgments based on tags,” and if a tag is wrong, they quickly run off course. My simple workaround is: when the alarm goes off, check permissions and balances once; trim positions when they need trimming; and don’t let the robot end up acting on my behalf at the wrong time.