Recently, I've been looking into AI agents running on-chain interactions, and I feel like we're still a breath away from "full automation." They can automatically quote prices, split routing, and even adjust positions according to rules, but when it comes to "who takes responsibility if something goes wrong," a human still needs to step in: for example, if an authorization accidentally grants unlimited access, or in cases of contract upgrades or fake front-ends—black swan events—no matter how smart the agent is, it's hard to rely on common sense to avoid trouble.

There are also issues like on-chain congestion, slippage fluctuations, and MEV front-running; if rules are hardcoded, they become easy targets for attacks.
In the past couple of days, I've also seen people tying ETF capital flows, US stock risk appetite, and crypto market rises and falls together, honestly, I’m too lazy to argue… sometimes data correlation is quite fragile, it’s satisfying to explain but painful to backtest.
Anyway, my current attitude towards agents is: let them do repetitive tasks, but the key buttons are still pressed by me. If they’re slow, so be it—at least I won’t wake up to an empty account in the middle of the night.