AI Agents can now click buttons on the chain by themselves, but honestly, I’m still not quite willing to fully let go... It can calculate paths, choose pools, but when slippage suddenly increases, gas fees spike, or a strange authorization gets inserted into the routing, I still have to manually intervene: first check transaction simulation, then see how much approval was given, and also watch out for those contract addresses that are "one step away from a black hole," AI might not even know the pain.

Recently, L2s are arguing again about TPS, fees, subsidies, and so on. I’m always a step behind; by the time I catch up, the debate has already gone through a round. For me, a more realistic concern is: during cross-chain or cross-layer operations, confirmation times, risk warnings for bridges, and the pitfalls of rollback on failure—if the Agent makes a mistake, I foot the bill... it’s quite hard to keep it all under control. Let it do the dirty work and handle the tedious tasks, while humans monitor the key steps, at least sleeping a bit easier.