ReAct (Reason and Act): how an AI agent works

Short answer: ReAct stands for Reason and Act. It is an approach in which an AI agent keeps making a short consideration (what is the next step?), then takes an action with a tool (a search, a system lookup, a calculation) and then checks what that produced. The loop repeats until the task is done.

A regular chatbot answers from what the model already knows. An agent working the ReAct way can fetch information along the way and adjust its plan. That is the difference between guessing and checking.

Where does ReAct come from?

The term comes from the research paper ReAct: Synergizing Reasoning and Acting in Language Models by Shunyu Yao and colleagues (Princeton and Google, 2022). They showed that language models perform better on question and decision tasks when reasoning and acting are combined, rather than only reasoning or only picking actions.

The same idea now sits inside many agent frameworks and AI assistants that are allowed to use tools. It is not always literally called ReAct there, but the loop is recognisable.

How does the loop work?

A ReAct agent goes through three steps each time:

1. Thought (Reason): the agent works out what is still needed. For example: "I don't know the opening hours, I need to look them up."
2. Action (Act): the agent calls a tool, such as a search function, a database or a calendar.
3. Observation: the agent reads the result and takes it into the next thought.

This continues until the agent knows enough to finish the task, or until a set limit is reached.

A practical example

Say you ask an agent which campaigns delivered the most last month.

• Thought: I need last month's figures.
• Action: fetch the campaign data from the ad account.
• Observation: five campaigns, with costs and conversions.
• Thought: I need to calculate cost per conversion for each campaign.
• Action: run the calculation.
• Observation: campaign B performs best.
• Answer: an overview with campaign B on top and the reasoning attached.

Every step can be followed. That makes it easier to see where it goes wrong when the outcome is off.

Why is this useful?

• Less making things up: the agent looks facts up instead of guessing them.
• Course correction: if an action returns something unexpected, the agent adjusts its plan.
• Traceable: thoughts and actions can be read back, which helps with testing and improving.

What should you watch out for?

• Permissions: only give an agent the tools and access it really needs. Have a person approve actions with consequences, such as sending or deleting something.
• Cost and time: every step is another model call. Long loops get slow and more expensive. Set a maximum number of steps.
• Tool quality: an agent is only as good as the information it retrieves. Messy data gives messy answers.
• Not the same as React: React is a JavaScript library for building websites. It has nothing to do with this.

Want to know more about what agents do in practice? Read about AI agents as well.

Frequently asked questions

What does ReAct mean in AI?

ReAct stands for Reason and Act. An AI agent alternates reasoning with actions through tools and checks the result after each action, until the task is done.

What is the difference between ReAct and a regular chatbot?

A regular chatbot answers from what the model already knows. A ReAct agent can look up information along the way, query systems and adjust its approach.

Is ReAct the same as React?

No. React is a JavaScript library for websites and apps. ReAct is an approach for AI agents.

Who came up with ReAct?

Shunyu Yao and colleagues from Princeton University and Google described the approach in 2022 in the paper "ReAct: Synergizing Reasoning and Acting in Language Models".