fixaiprompt
All techniques
Glossary · Technique

ReAct (Reason + Act)

Also known as: ReAct prompting, Thought-Action-Observation

Alternate reasoning and acting in a tight loop. The dominant pattern for tool-using agents — think, act, observe, repeat.

When to use it

  • Multi-step agentic tasks that need to use tools.
  • Research / scraping / search workflows.
  • Anywhere the model needs to react to real-world feedback before its next step.
  • When pure CoT would hallucinate facts it doesn't have.

When not to use it

  • Tasks that fit in a single shot — overhead isn't worth it.
  • When tool outputs are huge — observations blow up context.
  • Closed-domain tasks with no need for external information.

How it works

  1. 1Each step has 3 parts: Thought (what to do next + why), Action (the tool call), Observation (the result).
  2. 2Loop continues until the model emits a final answer.
  3. 3Forces the model to interleave reasoning with actions rather than deciding all moves up front.
  4. 4Critical: structure the prompt with explicit Thought:/Action:/Observation: markers.

Example

Lazy prompt
Find me the top 3 restaurants in Paris and book a table.
Using the technique
You are a ReAct agent. For each step output:
Thought: <reasoning>
Action: <tool_name>(<args>)
Observation: <returned by tool>

Repeat until you have a final answer, then output:
Final: <answer>

Max 8 steps. If stuck, output Final with what you have.

Common pitfalls

  • Models can fall into action loops (calling the same tool repeatedly without progress).
  • Long loops accumulate huge context — must summarize or truncate observations.
  • Tool errors must be caught and re-injected as observations, or the loop derails.
  • Without a max-step limit, the loop can spiral.

Where this came from

Yao et al., 2022 — "ReAct: Synergizing Reasoning and Acting in Language Models". Foundation of LangChain's early agent design.

Related techniques

Function Calling / Tool Use

Let the model decide when to invoke a real function or API instead of free-text answering. The foundation of every modern agent.

Retrieval-Augmented Generation (RAG)

Don't train on it — retrieve it. Inject relevant documents into the prompt at runtime so the model answers from real source material.

Chain-of-Thought (CoT) Prompting

Force the model to think step-by-step before answering. Dramatically improves accuracy on multi-step problems.