ReWOO Agent Example

Why it matters

ReWOO trades the ability to react to each observation for a predictable tool plan. Running it next to a ReAct agent shows when that trade pays and when it costs you a correct answer.

What to practise

Run the rewoo_agent workflow with nat run on the README’s question — who would be older today, Einstein or Bohr — with the calculator and internet_search tools available, then score it with nat eval. Follow the three nodes: the planner writes the whole plan, the executor runs tool calls that do not depend on each other in parallel, level by level, and the solver composes the answer. Read the two options that shape failure, tool_call_max_retries and raise_tool_call_error, and the planner_prompt and solver_prompt overrides.

How it fits

This is the plan-then-execute side of 1.2.3: compare it with ReAct, which decides each action after seeing the previous observation. Set raise_tool_call_error against 1.2.2’s question of how a tool failure should end, or not end, a run.