ReWOO cuts AI agent model calls to two regardless of task complexity

ReWOO, short for Reasoning WithOut Observation, is an AI agent framework that reduces large language model calls to just two per task, no matter how many reasoning steps are required. Unlike the popular ReAct approach, which triggers a new model call at every thinking step and re-sends the growing transcript each time, ReWOO separates the process into a Planner, a Worker, and a Solver. The Planner makes one model call to map out the entire tool-use plan upfront using variable placeholders, while the Worker executes each tool call sequentially in plain code without involving the model. Once all evidence is gathered, the Solver makes a single final model call to produce the answer. This design keeps model-call count constant regardless of task length, significantly reducing both API costs and token usage compared to ReAct on multi-hop reasoning tasks.