ReAct

ReAct (Reasoning + Acting) is the foundational AI agent pattern from Yao et al. (2022). The agent runs a loop: think about the problem, act through a tool, observe the result, then think again with the new information. Instead of one-shot generation, the agent iterates until the task is done — which lets it handle tasks that need outside information (search, calculation, API calls). Most modern agent frameworks (LangChain, AutoGen, and others) default to a ReAct-style loop; the main alternative is Plan-and-Act.