These days I’ve been watching AI Agents interact with the blockchain—it’s pretty lively to talk about. But once it actually runs in practice, I find myself caring more about one thing: which steps still need humans to cover. For example, with authorization—fine if you can’t read the contract, but the permission scope, whether it allows unlimited authorization, and whether the logic can be upgraded and changed: those “last look” checks, I still don’t dare hand over to the model. And with cross-chain stuff, like switching routes—Agents are great at finding the optimal path, but once you land in unfamiliar pools or a new bridge, who’s going to take the blame for the probability of slippage or getting stuck? Basically, when something goes wrong, you need to be able to trace back to which step went wrong.

As for the whole staking and shared security setup—the stacked returns getting criticized as a “nested doll” lately—I totally get that too. If an Agent only focuses on the annualized numbers, it might end up covering up risk layer by layer, and in the end push you into a bunch of places that are highly correlated. Anyway, my current approach is: automation is fine, but the key signatures, the limits/quotas, and the exit path still need to be confirmed by yourself—if it takes longer, then it takes longer. …Off to work.