Recently, AI agents have been running on-chain “doing their own work.” In plain terms, it can save a lot—points upon points—but when it actually comes to implementation, you still need someone to cover the risks. For example, in the authorization step: if the agent spots a cheaper path, it might approve with an unlimited allowance. As someone who loves to treat a pool like a library—organizing and filing things—I still manually cap the allowance and distribute funds in batches. Another example is routing/slippage: the agent only looks at quotes, but when liquidity is thinner, and MEV squeezes, everything gets distorted. At that point, people have to own the boundary of “it’s okay not to execute a trade.” There’s also contract upgrades and proxy address changes—no matter how smart the agent is, it still fears being tricked into signing by phishing pages. That’s why people need to take another look at the source and the permissions.

I also happened to think about the recent controversies around NFT royalties; they’re actually similar. Everyone wants to automate and hardcode creator earnings, but when secondary liquidity tightens up, the market “votes with its feet.” Later, I found it kind of ridiculous: on one hand, we want the machine to be more “automatic,” and on the other, we keep adding “manual confirmation” as a safety valve… Anyway, my current approach is: the agent handles finding routes and calculating things, and I only handle the final signature and the risk-control button. For now, that’s it.