Lately, I've been looking at AI agents interacting on-chain, and I feel they're still a bit short of being "fully automated." They handle checking balances, price comparisons, and splitting transactions quite smoothly, but when it comes to signing, I still prefer to keep an eye on it myself: authorization scope, whether the approval is unlimited, whether it's connected to a fake contract with the same name... In other words, humans need to be the final brake.

And then there's cross-chain operations, parameter modifications, emergency liquidity withdrawals—agents might calculate quickly, but when faced with congestion, rollbacks, or MEV front-running, their strategies tend to go off course more easily.

Now, with new L1/L2s offering incentives to attract TVL, it's normal for old users to complain about "mining, dumping, and selling." The agent sees profits and rushes in—who's managing slippage and fees when exiting?

Anyway, I prefer to let it be an assistant, not a driver.