ReAct Framework
ReAct (Reasoning + Acting) is a prompting paradigm and behavioral framework for llm-based agents. Unlike standard Chatbots that only generate text, ReAct enables agents to reason through complex tasks by interleaving thought processes with external actions.
Core Mechanism
The framework operates on a loop of three steps:
- Thought: The model reasons about the current state and decides on a next action.
- Action: The model executes an operation (e.g., calling an application-programming-interface-api, searching a knowledge-base).
- Observation: The model processes the output of the action, feeding it back into the reasoning loop.
This iterative cycle allows agents to handle multi-step problems, verify information, and adapt to dynamic environments.
Integration & Context
- Distinction from Chatbots: ReAct defines AI agents by their capacity to act and adapt, moving beyond static pattern matching AI Agents Explained: ReAct Framework, Behavioral Types, and Google ADK.
- Google ADK: The principles of ReAct are foundational to modern agent development kits, including Google ADK, which structures how agents build behavior types and manage tool use.
References
AI Agents Explained: ReAct Framework, Behavioral Types, and Google ADK