ReAct (Reasoning + Acting) - Reasoning + Acting Paradigm

Basic Information

• Type: AI Agent Reasoning Paradigm
• Paper: "ReAct: Synergizing Reasoning and Acting in Language Models" (2023)
• Authors: Shunyu Yao et al. (Google Research + Princeton University)
• Publication: ICLR 2023
• Current Status: De facto standard for agent AI frameworks

Paradigm Description

ReAct is a machine learning paradigm that unifies reasoning and acting. It allows LLMs to alternately generate reasoning traces and task-specific actions. Reasoning traces help the model deduce, track, and update action plans while handling exceptions, while actions enable the model to interact with external resources (e.g., knowledge bases or environments) to acquire additional information. ReAct has been adopted as a pre-configured agent module by mainstream frameworks such as LangChain, LangGraph, BeeAI, and LlamaIndex.

Core Mechanism

• Alternating Reasoning and Acting: Reasoning traces and task actions are generated alternately
• Reasoning Traces: Help the model plan, track progress, and handle exceptions
• Action Execution: Interact with the external environment to gather information
• Observation Integration: Integrate environmental feedback into subsequent reasoning
• Iterative Loop: Continuous cycle of thinking → acting → observing

Technological Evolution (2025-2026)

• Focused ReAct: Presets the original query at each step to maintain focus, improving accuracy by +530% and reducing runtime by -34%
• Multimodal ReAct: Integrates text, image, and video data, with LLMs invoking external visual experts via prompt encoding
• Industry Standard: Adopted as the de facto standard by LangChain and LangGraph
• Framework Integration: BeeAI, LlamaIndex, and others provide pre-configured ReAct agent modules

Performance

• Significantly outperforms Chain-of-Thought or action-only methods in tasks such as decision-making (AlfWorld) and reasoning Q&A (HotPotQA)
• Enhances interpretability, robustness, and generalization capabilities
• The Focused ReAct variant achieves over 5x improvement in accuracy

Relationship with Other Paradigms

• vs Chain-of-Thought: CoT only reasons without acting, while ReAct combines reasoning with acting
• vs Act-Only: Action-only methods lack reasoning capabilities, whereas ReAct adds planning and reflection
• vs Reflexion: Reflexion builds on ReAct by adding self-reflection and memory

Relationship with the OpenClaw Ecosystem

The ReAct paradigm is the core reasoning framework for OpenClaw's personal AI agent. Agents should adopt the "think → act → observe" loop when executing tasks, ensuring that each action is supported by reasoning and each observation updates subsequent planning. OpenClaw can use ReAct as the default agent reasoning strategy and integrate improved variants such as Focused ReAct on top of it.