In plain terms
A mechanic facing a car that will not start works in small rounds. She forms a thought (“probably the battery”), tests it (measures the voltage), reads the result (“the battery is fine”) and thinks again (“then the starter”). Thinking alone would leave her guessing; swapping parts alone would waste the afternoon. ReAct, short for reason and act, has a model work in the same rhythm: thought, action, observation, next thought.
Why it matters
Nearly every agent in use today descends from this idea, so the term is worth recognising: when a supplier says “ReAct agent”, it means the basic loop of thinking, using a tool and looking at the result, and nothing more exotic. Its value is that each thought is checked against a fresh fact before the next one is formed. Its limits follow from the same design: the agent decides one step at a time and can wander on long tasks, every step is a paid model call, and the written thoughts are no guarantee of the real reasons.
Example
An analyst asks which of two packaging suppliers has the lower defect rate this year. Thought: I need both rates. Action: query the quality database for supplier A; observation: 1.8%. Thought: now supplier B. Action: the same query; observation: no entries after March. Thought: the data is incomplete, try the inspection log. Action: search the log; observation: 2.6%. Answer: supplier A, with a note on where B's figure came from.
Most often confused with
ReAct vs. Chain of Thought (CoT)
Chain of thought makes a model reason step by step inside one reply; everything it reasons about was in the prompt or in its training. ReAct adds the outside world: between thoughts the model looks something up or changes something, and the next thought starts from what came back. Use chain of thought for a problem that can be solved on paper, and ReAct when the facts have to be fetched.
Origin: Introduced in 2022 by Shunyu Yao and colleagues at Princeton University and Google in the paper “ReAct: Synergizing Reasoning and Acting in Language Models”.
Under the hood
The 2022 paper prompted a model with a few worked examples in a fixed text format (Thought, Action, Observation), parsed each action in code, ran it and appended the observation. The method was tested on question answering (HotpotQA), fact verification (FEVER), a text game (ALFWorld) and web shopping (WebShop), and compared with reasoning alone and acting alone. Today the mechanics have changed and the pattern has stayed: native tool calling replaces text parsing, and reasoning models think between tool calls without any prompt format. Many frameworks still label their default agent a “ReAct agent”. Alternatives and extensions: plan-and-execute, which writes a full plan before acting, and Reflexion, which adds a written self-critique after a failed attempt. Pitfalls: repeating an action that keeps failing, long traces that fill the context, and observations that carry injected instructions. The name has no connection with React, the JavaScript library for user interfaces.