Papers
ReAct: Synergizing Reasoning and Acting in Language Models
Yao et al., 2022 — The paper that shows how LLMs solve complex tasks by alternating between thinking and acting.
ReAct (Reasoning + Acting) is a prompting pattern where an LLM alternates between thinking (reasoning) and doing (acting). Instead of just generating text, the model can call tools, interpret results, and adapt its approach. ReAct is the foundation for most current AI agent frameworks.
The Problem: Thinking Alone Isn't Enough
Chain-of-Thought (CoT) prompting showed that LLMs improve when they write out their reasoning step by step. But pure thinking has limits: The model cannot access current information, perform calculations, or query external systems.
Conversely, there are systems that let LLMs use tools (acting), but without explicit reasoning. These often act blindly — without a plan, error analysis, or strategy adjustment.
The ReAct Idea: Thinking AND Acting
ReAct combines both in an alternating loop:
- Thought: The model thinks — analyzes the current situation, plans the next step, interprets previous results.
- Action: The model executes a concrete action — e.g., a web search, calculation, or API call.
- Observation: The result of the action is returned to the model. It serves as input for the next thought.
This cycle repeats until the task is solved.
ReAct in Action: An Example
Question: "In what year was the capital of the country where the Transformer was invented founded?"
Why ReAct Works Better
- Transparency: The thought steps make the reasoning process traceable. You can see why the model chose a particular action.
- Error correction: When an action returns an unexpected result, the model can adjust its approach instead of stubbornly sticking to a wrong strategy.
- Grounding: Through actions (search, computation), answers are based on real data rather than hallucinations.
- Flexibility: The pattern works with any tools — web search, databases, APIs, code execution.
ReAct in Today's Agent Frameworks
ReAct is the basis of virtually all AI agent frameworks today:
- LangChain Agents: Implement the ReAct loop as the default agent type
- Claude Tool Use: Anthropic's function calling follows the Thought-Action-Observation pattern
- AutoGPT / CrewAI: Multi-agent systems where each agent internally uses ReAct
- Claude Code: Also uses the ReAct pattern for code analysis and generation
Sources
- Yao, S. et al. (2022). "ReAct: Synergizing Reasoning and Acting in Language Models." arXiv:2210.03629
Related articles
Agent Orchestration Patterns
Proven orchestration patterns for Multi-Agent Systems: sequential, parallel, hierarchical, router, and supervisor.
Retrieval-Augmented Generation (RAG) Paper Explained
The RAG paper by Lewis et al. (2020) explained: How LLMs become better and more reliable through external knowledge sources.
Constitutional AI Explained
The Constitutional AI paper by Bai et al. (2022, Anthropic) explained: How to align AI systems through principles instead of human feedback alone.
Related articles
Agent Orchestration Patterns
Proven orchestration patterns for Multi-Agent Systems: sequential, parallel, hierarchical, router, and supervisor.
Retrieval-Augmented Generation (RAG) Paper Explained
The RAG paper by Lewis et al. (2020) explained: How LLMs become better and more reliable through external knowledge sources.
Constitutional AI Explained
The Constitutional AI paper by Bai et al. (2022, Anthropic) explained: How to align AI systems through principles instead of human feedback alone.
Was this article helpful?
Continue the learning path
The learning path puts these articles in order, and the Hub carries the building blocks we have checked in our own operations.
- Local and self-hosted
- Documented and verifiable
- From our own operations
- Made in Austria