Lately I've been seeing a bunch of AI agents claiming they can automatically go on-chain and do work, and I’m half convinced. Checking balances, calculating routes, even monitoring the mempool to choose a better timing for sending transactions—these are fine for them to run; but when it comes to the actual "signing" step, humans still need to back it up. Honestly, once you give them authorization, they can treat your wallet like a self-service buffet.

There are also a few things I think must be handled manually: the whitelist for new contracts/new routes (at least to check if it’s freshly deployed, if the code looks copied), don’t set the approval limit all at once, and don’t blindly push on retry strategies. When the chain is congested, agents can easily turn into “mindless accelerators,” spending more gas without necessarily being faster.

As for staking and shared security setups, which have recently been criticized as “cloning,” agents need to be more cautious: stacking yields looks attractive, but the actual risks are layered and transmitted. No matter how smart the model is, it can’t read “human nature” or the bottom line of project teams. Anyway, my current approach is: automate what can be automated, don’t sign if you don’t have to, and if you really need to sign, do a small test first to feel more at ease.