Recently, seeing AI Agents running on-chain to do their own tasks has been quite exciting and a bit nerve-wracking. To put it simply, it can skip the step of "pressing buttons," but if something really goes wrong, humans still need to take responsibility: who was given the signing authority, whether the contract is just a shell swap, whether slippage and routing have been tampered with, whether cross-chain steps are stuck on the bridge... it can calculate these, but it doesn't "fear" them; humans do, so humans need to keep an eye on it.

Another thing is that when narratives heat up, memes and celebrities shouting out grab attention, and the focus shifts. I admit I also envy others who hit the jackpot by going all-in on hot trends, but honestly, the hardest part is not to take the final baton. If an Agent is fed a "chasing hot topics" objective function, it might get even more obsessive than humans. Anyway, I now prefer to let it handle execution, while I set my stop-loss lines and permission boundaries. Even if it’s slower, that’s fine.