ReAct vs Planning: when your agent stops making progress

An engineering team had been debugging the same agent for three weeks. It would search, then search again, then search a third time, then output a half-finished answer at the iteration limit. The model was fine. The pattern was wrong. Here is the dif