ReAct (Reasoning + Acting) is a prompting and training framework for language models that interleaves verbal reasoning traces with discrete actions, allowing a model to plan, execute tool calls, observe results, and revise its reasoning within a single loop. Published by Yao et al. in 2022, it formalises the intuition that thinking and doing should alternate rather than happen in isolation.

Think of it like a chef who pauses mid-recipe to taste, adjusts the seasoning based on what they notice, then continues cooking. Rather than committing to a full plan upfront and executing it blindly, the chef reasons, acts, observes, reasons again-each step informed by the last.

A ReAct agent answering "What is the population of the capital of France?" might produce: Thought: I need to identify France's capital, then look up its population. Action: Search("capital of France"). Observation: Paris. Thought: Now search for Paris's population. Action: Search("Paris population 2024"). Observation: ~2.1 million. Answer: Approximately 2.1 million. This explicit trace makes failures easy to diagnose and correct.

ReAct sits at the intersection of Chain of Thought reasoning and tool-augmented Inference; it is commonly paired with RAG pipelines and Prompt Engineering techniques that structure the thought/action/observation cycle. The framework influenced later agent architectures such as Toolformer and many modern LLM agent libraries.

You will encounter ReAct whenever you use LLM-powered agents that call external tools-web search, code interpreters, or APIs-and need an auditable reasoning log alongside each action.