Lately, watching AI agents interact on the blockchain has been pretty exciting, but when it really runs, someone still needs to take responsibility. The most annoying part is authorization; agents give unlimited credit to save gas, but when they encounter phishing contracts, they get wiped out instantly; routing/slippage issues are the same—sudden queue jumps on the chain, MEV snipes, and the agent just follows the strategy, while people need to monitor for abnormal transactions. Also, with oracle price feeds and liquidation thresholds, models don’t understand “boundary conditions”—if the data source jitters, it can trigger a chain reaction, and in the end, someone still has to manually shut down or switch sources.

Airdrop season now uses a points system plus anti-witchcraft measures; agents automatically do tasks, which is like clocking in at work… but honestly, they can’t take responsibility for risk control and identity verification. My approach is pretty simple: small amounts, separate wallets, permissions on demand, strategies can be handed over to them, but the consequences shouldn’t be.