ReWOO (Reasoning WithOut Observation)
ReWOO is an agent architecture that separates planning from execution. The model produces the full plan up front — including every tool call with placeholder variables for tool outputs — before any tool runs. Tools then execute in sequence or in parallel, and a final solver pass integrates the actual results into the final answer.
This removes the interleaved reasoning-action-observation loop that ReAct uses, cutting token cost because the model does not re-read and re-reason over each intermediate observation. The trade-off is adaptivity: if a tool returns something the planner did not anticipate, ReWOO cannot easily change course mid-execution, whereas ReAct can. ReWOO suits well-structured tasks with predictable tool behavior.
Example
A travel agent receives "Plan a 3-day trip to Kyoto within $1,200". The planner emits a single plan: `t1 = search_flights(origin, "KIX", date_range)`, `t2 = search_hotels("Kyoto", nights=3)`, `t3 = suggest_itinerary(nights=3, interests=[#E1], hotel=#t2)`. Tools run, then the solver composes an itinerary referencing the actual flight and hotel prices. Compared to a ReAct loop that re-prompts the model after each tool call, ReWOO saves several thousand tokens on the same task.
Frequently asked questions
What is ReWOO (Reasoning WithOut Observation)?
- ReWOO is an agent architecture that separates planning from execution. The model produces the full plan up front — including every tool call with placeholder variables for tool outputs — before any tool runs.
How does ReWOO (Reasoning WithOut Observation) work?
- This removes the interleaved reasoning-action-observation loop that ReAct uses, cutting token cost because the model does not re-read and re-reason over each intermediate observation. The trade-off is adaptivity: if a tool returns something the planner did not anticipate, ReWOO cannot easily change course mid-execution, whereas ReAct can.
Can you give an example of ReWOO (Reasoning WithOut Observation)?
- A travel agent receives "Plan a 3-day trip to Kyoto within $1,200". The planner emits a single plan: `t1 = search_flights(origin, "KIX", date_range)`, `t2 = search_hotels("Kyoto", nights=3)`, `t3 = suggest_itinerary(nights=3, interests=[#E1], hotel=#t2)`. Tools run, then the solver composes an itinerary referencing the actual flight and hotel prices. Compared to a ReAct loop that re-prompts the model after each tool call, ReWOO saves several thousand tokens on the same task.
Put this into practice
Build polished, copy-ready prompts in under 60 seconds with SurePrompts.
Try SurePrompts