PAPER 14 · Agents

ReAct: Synergizing Reasoning and Acting

Yao et al. 2022 / ICLR 2023 Paper

The foundation of practical agent loops: reasoning traces combined with tool use and environment interaction.

Core concept

ReAct gives a model a loop: think, use a tool, observe the result, then think again.

Why it mattered

It is one of the roots of modern agents and tool-using LLM workflows.

Visual shortcut · Agent loop
reason
act
observe
reasoning gets corrected by the world

ReAct is the basic agent rhythm: reasoning gets stronger when the model can check the world and update.

How it works
Write a thought about what is needed.
Call a tool or take an action.
Read the observation.
Continue with new information.

The quick digest

A normal model answer is one shot: prompt in, text out. ReAct changes the rhythm. The model writes a thought, chooses an action like search or lookup, reads the observation, and continues from there.

That matters because many tasks cannot be solved from memory alone. The model may need current information, calculation, environment feedback, or a chance to correct a bad assumption. ReAct gives the model a structured way to combine internal reasoning with external evidence.

Modern agent systems are more engineered than the paper’s format, but the core loop remains: reasoning is stronger when it can act, observe, and update.

What to remember

One-liner
Agents are loops: think, act, observe.
Why it matters
Tools let the model update from the world.
Builder instinct
The loop matters more than the theatrics around it.

Read it like this

Build instinct

Make a two-tool agent that searches notes, performs one calculation, and explains when each observation changed its answer.

Read source → All papers
Previous13 · Chain-of-Thought Prompting Elicits Reasoning in Large Language Models