Agent Trajectory Schema

A new analysis argues that reinforcement learning from human feedback (RLHF) is insufficient to ensure the safety of autonomous AI agents, advocating instead for runtime contracts that enforce hard bo…