ReAct (Reasoning + Acting)
A prompting paradigm that connects reasoning (thinking) and acting (doing) in a loop – the LLM thinks aloud, executes actions, and reflects on results.
ReAct connects thinking and acting: Think → Act → Observe → Repeat. The base pattern for robust AI agents.
Explanation
ReAct agents follow a thought-action-observation loop: Thought analyzes the situation, Action executes a tool, Observation processes the result. Repeat until goal reached. Much more robust than pure chain-of-thought as errors in intermediate steps can be corrected.
Marketing Relevance
Standard pattern for agentic AI. LangChain, AutoGen, and other frameworks implement ReAct as the base architecture for tool-using agents.
Example
Thought: "I need current weather data." → Action: weather_tool(Berlin) → Observation: "15°C, sunny" → Thought: "Now I can formulate the recommendation." → Final Answer.
Common Pitfalls
Token-intensive due to verbose reasoning. Can get stuck in loops. Requires good prompt engineering for consistent format.
Origin & History
ReAct was introduced in 2022 by Yao et al. (Princeton, Google) in the paper "ReAct: Synergizing Reasoning and Acting in Language Models." It was the first to combine CoT reasoning with tool use.
Comparisons & Differences
ReAct (Reasoning + Acting) vs. Chain-of-Thought
Chain-of-thought only thinks; ReAct additionally executes actions and processes their results.
ReAct (Reasoning + Acting) vs. Plan-and-Execute
Plan-and-execute plans everything upfront; ReAct plans and corrects iteratively based on observations.