What is ReAct? — AI Glossary

What it is: ReAct is a design pattern for AI agents where the model alternates between Reasoning (thinking through the problem) and Acting (calling tools or taking actions in the world). It’s the foundation of how most AI agents in 2026 actually work under the hood.
Who it is for: Developers building AI agents, and anyone curious about how chatbots actually use tools like web search, calculators, or APIs.
Best if: You’re building an agent that needs to use tools, or trying to understand why some AI products feel more capable than others.
Skip if: You’re a casual user and just want to know what AI can do, not how the loop is structured. Want one practical AI workflow every morning? Subscribe to our free daily newsletter.

What is ReAct?

ReAct (Reason + Act) is an AI agent design pattern introduced in a 2022 Google paper. It structures the agent’s behavior as a loop alternating between two steps:

• Reason — the model thinks through the current state of the task, what it knows, and what to do next.
• Act — the model executes one action, typically calling a tool (web search, calculator, code interpreter, database query, etc.).
• Then observe the result, and start the loop again.

That tight loop — think, act, observe, think, act, observe — is how most modern AI agents actually work. The pattern combines the reasoning ability of large language models with the real-world capabilities of tool use, producing systems that can complete multi-step tasks no single model call could handle.

Why does ReAct matter?

ReAct is the conceptual backbone of the agent products that took off in 2024-2026 — ChatGPT agent mode, Claude with computer use, Gemini’s Project Mariner, Manus, Replit Agent. They all implement variations of the reason-act-observe loop. Understanding ReAct is the cleanest way to understand why agents can do things a single LLM call cannot.

The pattern also explains agent failure modes. Long ReAct chains accumulate errors: the model might reason wrong, observe ambiguous results, and propagate the mistake forward. Modern agent products tackle this with planning steps, verification steps, and human-in-the-loop checkpoints. But the basic loop is unchanged.

How do you use ReAct in practice?

For users: you don’t implement ReAct directly — it’s baked into the agent products. But understanding the pattern helps you write better agent prompts. Effective patterns:

• Be explicit about goals and constraints (the agent’s reasoning happens against your stated objective)
• Allow checkpoints (“After researching, summarize what you found before continuing”)
• Give it the tools it needs (web access, file uploads, code interpreter)
• Limit scope (ReAct loops drift when goals are too open-ended)

For developers: the major LLM frameworks (LangChain, LangGraph, OpenAI Agents SDK, Anthropic’s SDK) all support ReAct-style agent construction. Start by defining the tools available, then let the model loop between reasoning and tool calls until it determines the task is complete.

Related terms

• AI Agent
• Tool Use
• Function Calling
• Chain of Thought
• Agentic Workflows

Learn more on Beginners in AI

• What Are AI Agents?
• AI Glossary

Sources and further reading

• ReAct: Synergizing Reasoning and Acting in Language Models (Google, 2022)

Last reviewed: May 2026. AI terminology evolves quickly — verify specifics on the official source pages above.

You may also like

• What Are AI Agents?
• AI Glossary
• AI Agent
• Tool Use
• Function Calling
• Chain of Thought