Lately, watching AI Agents run on the blockchain, I really get a sense of "it works really well but also causes a lot of trouble." It’s much calmer than me when placing orders, routing, calculating gas, but when it comes to signing, I still keep an eye on it manually: what exactly was authorized, is it an unlimited limit, has the contract address been swapped... Basically, the safety net isn’t in execution but in “what consequences are you willing to bear.”

And then, when the network suddenly gets congested, transactions get stuck, or cross-chain transactions don’t get a response for a long time, the Agent just keeps retrying. I tend to stop and decide whether to cancel, change the route, or even just give up. Especially now, with staking and shared security systems being criticized as “nested,” the compounded yields look tempting, but the risks also stack up—Agent won’t worry about that for you.

What I’ve learned isn’t techniques, but: it’s okay to treat the machine as an accelerator, but don’t let go of the steering wheel.