ReAct (Reason + Act)

What it is

ReAct ("Reasoning and Acting") is a prompting pattern that has the model alternate between generating reasoning ("thoughts") and taking actions such as calling a search tool or an API, then observing the returned result before continuing. Yao et al. showed that combining reasoning with acting outperforms either alone on knowledge-intensive and decision-making tasks.

By grounding each reasoning step in fresh observations from the environment, ReAct lets the model plan, gather missing information, correct course, and avoid the closed-loop hallucination that pure chain-of-thought can suffer when it lacks facts. It is the conceptual backbone of many modern tool-using agents.

When to use

• The task needs information the model does not hold: live data, private documents, calculations.
• You are building a tool-using or agentic workflow (search, code execution, APIs).
• The model must plan, act, observe, and adapt over several steps.
• You want reasoning grounded in verifiable observations rather than the model's memory alone.

How it works

1. Define the available tools/actions and the format for invoking them.
2. Prompt the model to emit a Thought (reasoning) then an Action (a tool call).
3. Execute the action and feed the Observation (result) back into the context.
4. Repeat Thought → Action → Observation until the model emits a final answer.

Illustrative structure

The structure is a repeating trace: "Thought: <reasoning about what to do next> / Action: <tool>[<input>] / Observation: <result>" — looping until the model outputs "Answer: <final response>". The scaffold defines the tool grammar; the model fills the thoughts and actions.

Pitfalls

• Requires an execution harness to run tools and inject observations — more infrastructure than plain prompting.
• Malformed action syntax breaks the loop; the format must be strict and validated.
• Loops can run away or repeat; enforce a step limit and stopping condition.
• Tool errors and noisy observations can derail reasoning if not handled gracefully.

Frequently asked questions

Sources

• Yao et al. (2022), ReAct: Synergizing Reasoning and Acting in Language Models
• Anthropic — Tool use (function calling)

Related techniques

• Chain-of-Thought Prompting
• Prompt Chaining
• Retrieval-Augmented Generation (RAG)