A loop where an agent alternates between thinking out loud and taking an action, over and over until the task is done.
ReAct stands for 'Reasoning and Acting.' It's a pattern for building agents: the model writes a THOUGHT (what it wants to do next and why), then an ACTION (the tool call it picks), then gets back an OBSERVATION (the tool result), then thinks again. This cycle continues until the goal is met. The interleaving of reasoning with acting makes the agent much more reliable than having it decide all steps upfront.
ReAct is the most common pattern under the hood of agentic systems. Claude's agent harness, LangChain's AgentExecutor, and most open-source agents implement some flavor of ReAct. Knowing the pattern helps you debug when your agent misbehaves - usually it's because the thought→action linkage broke somewhere.