ReAct (Reasoning and Acting)

Overview

ReAct, proposed by Google in 2022, has the LLM cycle through 'Thought → Action → Observation' steps, using external tools (search engines, calculators, APIs) to gather information and verify intermediate results. Most modern AI-agent frameworks are built on this pattern.

Relationship to AI agents

Frameworks like LangChain, LlamaIndex, OpenClaw, and Claude Code implement the ReAct pattern at their core. The think-act loop enables LLMs to handle tasks requiring information retrieval or computation that the model cannot perform in a single inference step.