The Idea

Most AI agent patterns work like this: think, use a tool, look at the result, think again, use another tool, look at that result, think again... Each cycle sends the entire growing conversation back to AI. The more steps, the more expensive it gets.

ReWOO (Reasoning Without Observation) takes a radically different approach: plan all the tool calls in one shot, using placeholders for results you don't have yet. Then run the tools without involving AI at all. Finally, give AI the collected results and ask for the final answer. Total AI calls: just two, regardless of how many tools you need.

Building Blocks

This composition improves upon:

Plan-and-Execute

ReWOO takes Plan-and-Execute's separation of planning and execution, and pushes it further — the execution phase uses zero AI calls, just tool execution with placeholder substitution.

See It in Action

Question: "What is the population of the hometown of the 2024 Australian Open winner?"

Planner (1 AI call)
1
AI plans all tool calls with placeholders
The plan
#E1 = Search "2024 Australian Open winner"
#E2 = Search "hometown of #E1"
#E3 = Search "population of #E2"

Notice: AI doesn't know the answers yet. It uses placeholders like #E1 to reference future results.

↓ no AI involved below
Worker (0 AI calls)
2
Tools run automatically, substituting results
Execute #E1
Search "2024 Australian Open winner" → Jannik Sinner
Execute #E2 (substitute #E1 = "Jannik Sinner")
Search "hometown of Jannik Sinner" → San Candido, Italy
Execute #E3 (substitute #E2 = "San Candido, Italy")
Search "population of San Candido, Italy" → 3,350
↓ give AI all the evidence
Solver (1 AI call)
3
AI synthesizes the final answer
Final answer
The 2024 Australian Open was won by Jannik Sinner, who is from San Candido, Italy. The population of San Candido is approximately 3,350.

The Cost Difference

Traditional Agent Pattern

Think → tool → think → tool → think → tool → think

4+ AI calls, each resending the full growing context. Cost grows with every step.

ReWOO

Plan (1 AI call) → tools run alone → Synthesize (1 AI call)

Exactly 2 AI calls, no matter how many tools. Uses roughly 5x fewer tokens.

Why This Works

The key insight is that AI doesn't need to see each tool result before planning the next tool call. If you ask "Who won the tournament, where are they from, and what's the population there?" — you already know the sequence of lookups needed, even without knowing the intermediate answers.

By planning with placeholders, AI does all its reasoning upfront in one shot. The tools execute mechanically, just substituting real values for placeholders. And the final synthesis has all the evidence it needs in one clean package.

The Composition

One AI call to plan all tool calls using placeholders. Tools execute on their own, filling in the blanks. One final AI call to synthesize everything. Two AI calls total, no matter how many steps.

When to Use This

When to Skip This

How It Relates

ReWOO is a cost-optimized evolution of Plan-and-Execute. Both separate planning from execution, but Plan-and-Execute still uses AI for each execution step. ReWOO eliminates AI from the execution phase entirely.

It's complementary with reactive agent approaches — for predictable multi-hop lookups, use ReWOO for efficiency. For unpredictable, exploratory tasks, use a reactive pattern. Some systems even combine both: start with ReWOO, and fall back to reactive mode when the plan breaks down.