Last night, I stared at the funding rate until my eyes went numb, and casually tried having the AI agent run a transaction on the blockchain for me... To be honest, it’s pretty good at “executing,” but when it comes to pitfalls, I still need to have a human backup. For example, the authorization/limit step—I’d rather click twice myself than let it open an unlimited authorization right away; and then there’s routing and slippage— the agent only looks at the best quote, it doesn’t have that kind of intuition about “this pool looks like it’s easy to get front-run.”

There’s also a more practical issue: on-chain data tools and tagging systems have recently been criticized for lagging behind and being easily misled, and the agent especially trusts tags. Once misled by fake tags, it’s as ridiculous as ordering takeout at midnight and ending up in a different city... Anyway, my current approach is: let it do the heavy lifting (checking large transfers, running multiple routes), but I still handle the final confirmation and signing myself. If I make a mistake, I admit it—don’t let the robot decide for me.