ReAct
ReAct是一种大语言模型驱动的Agent设计范式，由Yao等研究者于2022年提出。其核心思想是将推理（Reasoning）与行动（Acting）交替进行，使模型在执行任务时能够生成思维链式推理步骤，并与外部工具或环境进行交互，从而动态调整决策。ReAct范式被广泛应用于问答、信息检索和任务规划等智能代理场景。
Core Facts
The ReAct framework forms a three-part cycle of Thought → Action → Observation, which differs from Chain-of-Thought (CoT) that reasons only within the model internally.
ReAct框架由普林斯顿大学和Google Brain团队在2022年的论文中提出
ReAct（Reasoning + Acting）由谷歌研究团队于2022年提出，将思维链推理与外部工具调用交织在一起
Timeline (last 90 days)
从零构建 Agent 能让开发者理解 ReAct 等推理范式的工作原理
ReAct 由 Yao 等人在 2022 年提出，其核心思想是将推理和行动交替进行
代理式架构基于 ReAct 范式相比单次推理生成代码具有更强的错误恢复能力，但代价是系统行为复杂度呈指数级增长，一次任务可能涉及数十次模型调用和工具交互
高级提示技术包括Tree-of-Thought（思维树）、Self-Consistency（自洽性采样）、ReAct（推理+行动）等，通过设计输入格式在不改变模型参数的情况下激活模型潜在推理能力
2023年Yao等人提出的ReAct范式让模型交替进行思考（Thought）和行动（Action），奠定了现代Agent的基本工作模式
ReAct范式由普林斯顿大学的Shunyu Yao等人在论文《ReAct: Synergizing Reasoning and Acting in Language Models》中提出，该论文于2022年10月发表并被ICLR 2023收录
AutoGPT、BabyAGI是ReAct范式的早期实践项目
Draw.io Skill 采用生成后自检的 ReAct（Reasoning + Acting）循环与自我验证范式，通过生成-验证-修正闭环提高一次交互的成功率
Agent框架慢于直接对话主要因为注入大量系统提示词工具定义、ReAct多轮推理和上下文累积导致输入token数增长
标准模式拥有完整工具集（文件读写、命令执行、Web搜索等），采用经典的ReAct（Reasoning + Acting）循环
7 more timeline events
All Facts (20)
The ReAct framework forms a three-part cycle of Thought → Action → Observation, which differs from Chain-of-Thought (CoT) that reasons only within the model internally.
90%VerifiedReAct框架由普林斯顿大学和Google Brain团队在2022年的论文中提出
90%VerifiedReAct（Reasoning + Acting）由谷歌研究团队于2022年提出，将思维链推理与外部工具调用交织在一起
90%VerifiedA typical Agent is composed of a cycle of 'Perception—Planning—Action—Reflection'.
85%VerifiedThe ReAct framework was published at NeurIPS in 2022 by Yao et al., demonstrating that interleaving natural language reasoning traces with action steps significantly outperforms using chain-of-thought alone or action sequences alone.
80%VerifiedThe ReAct (Reasoning + Acting) framework was proposed by Yao et al. in 2022, alternating Chain-of-Thought reasoning with external tool interactions.
80%VerifiedAI Agent通常采用ReAct（Reasoning + Acting）模式，在推理和行动之间循环迭代
80%VerifiedThe ReAct pattern was formally proposed by Yao et al. in the 2022 paper "ReAct: Synergizing Reasoning and Acting in Language Models".
80%VerifiedReAct（Reasoning + Acting）是2022年由Google提出的Agent范式，核心思想是让大模型在推理和行动之间交替迭代
80%VerifiedCursor通过对项目文件建立向量索引来理解整个代码库的上下文结构
80%VerifiedClaude Code's Agent architecture is based on the ReAct framework (Reasoning + Acting, an alternating cycle of reasoning and action).
80%VerifiedLLM-based agent frameworks typically adopt design patterns such as ReAct (Reasoning + Acting) or Plan-and-Execute.
70%Verified智能体循环允许AI模型在一次任务中反复经历感知、规划、行动、观察的迭代循环，并将函数返回值纳入下一轮推理上下文
65%Verified自主Agent基于ReAct（Reasoning + Acting）框架进行'感知-规划-行动-观察'循环，对初始上下文质量要求更高
65%VerifiedReAct要求模型在每一步明确输出「思考—行动—观测」的三元组，形成可追踪的决策轨迹
65%UnverifiedA typical Agent execution loop includes perception, reasoning, action, and observation, forming the ReAct (Reasoning + Acting) pattern
85%UnverifiedReAct论文发表于2022年10月
85%UnverifiedCodeAct实验数据显示其成功率显著高于JSON-based Action（ReAct模式）
75%UnverifiedReAct和Plan-and-Execute等主流Agent框架都可以将MCP作为工具调用层的标准化实现
75%UnverifiedOpenAI's Function Calling, Anthropic's Tool Use, and the ReAct framework have moved LLM-driven agents from academic concepts into engineering practice.
60%