ReWOO — Planner, Workers, Solver (Xu et al., 2023)
Planner
one call; sees only
the user question
plan DAG and workers
E1 = search
capital of France
E2 = search
population of #E1
E3 = round_million
text = #E2
evidence table: E1=Paris, E2="11.2M metro", E3="11 million"
Solver
one call; reads
evidence + plan
why ReWOO
no observation in planner
means no feedback noise
workers fan out in parallel
along independent branches
fewer tokens:
1 planner + N small worker
prompts + 1 solver call
paper: 5x reduction
on HotpotQA, +4 accuracy
when to pick it
task shape is structured
tools are known up front
evidence is parallelizable
token budget matters
else prefer ReAct for
unknown environments or
Plan-and-Act for 30+ step
web/mobile trajectories
distillation angle (ReWOO section 4)
planner sees no observations, so its training signal is clean; LLaMA 7B can be fine-tuned from
GPT-3.5 planner traces and match planner quality. 2026 production: small planner + large executor
(or vice versa) via model router (Mastra) or dynamic handoff (OpenAI Agents SDK).