ReAct vs ReWOO: Which Agent Execution Pattern Fits?
Use ReAct when the next action depends on a tool result you cannot predict in advance. Consider ReWOO when a model can describe the needed steps and evidence dependencies before execution. A planner-executor design helps when an explicit plan needs controlled execution and revision.
Choose the pattern from the task’s dependencies and recovery requirements, then measure its model calls and mistakes.
When does the model inspect tool results?
| Pattern | Decisions | Practical trade-off |
|---|---|---|
| ReAct | Model chooses an action after reading observations | Adapts to new evidence, with repeated model calls |
| ReWOO | Planner writes steps and evidence references; solver receives results | Less repeated planning context, with dependencies resolved during execution |
| Planner-executor | Planner creates a plan; execution may trigger replanning | Explicit progress and recovery, with extra plan state |
The ReAct paper interleaves reasoning, actions, and observations. This fits exploratory search where each result determines the next query. Retaining every prior result can increase input length; context management remains part of the implementation.
ReWOO separates planning, tool work, and final synthesis. Its steps can refer to earlier evidence: first obtain a location, then search for restaurants near that location. The second call must wait for the first. Only independent steps can run in parallel.
Specify recovery before comparing cost
A plan needs a policy for a missing tool result, invalid arguments, and a result that changes the task. Decide whether execution retries, stops, asks for clarification, or returns to the planner. A fixed plan without that policy can fail even when its original steps looked reasonable.
Run the same representative tasks through the candidate patterns. Count successful tasks, wrong actions, model calls, retries, tokens, and elapsed time. Include dependent steps and tool failures; an all-success demonstration cannot show recovery behavior.
There is no universal token-cost order among these patterns. Task length, tool output size, retries, and implementation can reverse the comparison.
For the earlier automation choice, see agent vs workflow. The agent-patterns section of the LLM Engineering Guide connects orchestration to tool calling and structured output.