Recently, I saw a bunch of AI Agents claiming to help you achieve "full automation" on-chain.

My first reaction was: Sure, but does that mean liquidation and theft are automatic too…?
Honestly, there are still quite a few steps that really need human oversight.

For example, during authorization, no matter how smart the Agent is, it might pretend to be naive and sign an unlimited limit, or change the interaction target to a phishing contract that looks similar;
And for cross-chain/bridges, if something goes wrong, it just shows "routing in progress," and you can only curse at the blockchain explorer.
Then there are parameters—slippage, gas, whether to use a private pool—Agents might choose the easiest route to save effort, but the easiest isn’t always the safest.

Recently, testnet incentives and point expectations have become like lottery tickets, everyone guessing whether the mainnet will issue tokens.
Agents might be more "diligent" than you in refreshing interactions, but spamming a bunch of "witch" tags is pretty amusing.
Anyway, my current approach is pretty simple: I review the key signatures myself, keep the limits small if possible, and if I really want to push, I start with small amounts to test and learn…
It's discouraging, but at least I don’t want to put all the blame on the robots.