Glossary
What is a ReAct agent
In short
A ReAct agent is an AI agent that works in turns: it reasons, calls a tool, looks at the result and reasons again.
AI agents · ReAct (Reasoning and Acting)
MMOOZZGG.com editorsChecked: 1 min read
For a child
Think of how you look for a lost ball. First you think: maybe it is under the sofa. You look there, but the ball is gone. Then you think again: maybe it stayed in the yard. You go out and find it by the fence. You switch between thinking and checking until you find it. A smart program can learn to look for an answer the same way: think, check, think again.
In plain words
The name joins two English words: reasoning and acting. The model does not give an answer at once. It writes a short thought about what it needs to find out. Then it takes an action, for example a search in a knowledge base or opening a web page. It gets the result, writes a new thought and decides what to do next. The loop repeats until the answer is found. Take the question “is this item in stock at the Kazan warehouse, and when will it arrive”. It takes several steps: first the stock, then the delivery schedule, then the answer to the customer.
For business
Many agents that work with company systems, such as the CRM, the warehouse and email, use this scheme. Its strength is transparency: the log of thoughts and actions shows why the agent made a decision, so an error is easy to find. The weak points are cost and speed. Each step of the loop is a new call to the model, so a long task costs much more than one answer and takes longer. If the agent gets stuck in a loop, spending grows with no result, so you set a step limit and a budget per task. Before rollout, test on real requests how many steps an average task takes and how often the agent picks the wrong tool.
For a specialist
ReAct (Reasoning and Acting) was proposed by Yao and co-authors in the paper “ReAct: Synergizing Reasoning and Acting in Language Models” (arXiv, 2022; the final version appeared at ICLR). The model produces alternating traces of reasoning (thought) and action, and the environment returns observations. Reasoning helps to build and correct a plan; action brings in outside information. On HotpotQA and FEVER, calling a simple Wikipedia API reduced hallucinations and the spread of errors, compared with chain-of-thought prompting. On ALFWorld and WebShop, the method beat imitation learning and reinforcement learning by 34 and 10 percentage points of success, with one or two examples in the prompt. Today the scheme is built with function calling. The thought stays as text or as hidden reasoning, and the action is a tool call. Reflexion extends the idea by adding a review of failed attempts.
Do not confuse. Not to be confused with React, the library for building web interfaces: only the letters of the name match.
Example
A customer asks in a chat whether delivery can move to Saturday. The agent reasons that it needs the order number and finds it in the CRM. It checks the free courier slots and only then offers a time.
Related terms
Questions and answers
How is ReAct different from chain-of-thought?
Sources
Not sure whether your company needs this?
AI solutions map