AI agents transacting freely, under control

AI agents will manage user wallets, but how do we prevent them from making human-like mistakes? We propose separating transaction construction from agent decisions, using a formally specified execution engine to ensure security and correctness.