ReAct agent
Reasoning + Acting — an agent pattern (paper 2022) where an LLM alternates reasoning ("Thought"), action ("Action", e.g. a tool call), and observation ("Observation") in a loop until the task is solved. The fundamental recipe behind most LLM agents.
Mechanic: the model produces a thought ("I need to look up X"), then an action (calls a tool), reads the observation (the result), and reasons on ("now I know X, next I need Y") — the loop repeats until it reaches an answer. Writing out the reasoning makes the decisions more reliable (like chain-of-thought), and being able to act between steps lets it fetch information and correct itself. Win: it combines reasoning with actual tool use → can solve multi-step tasks that pure prompting can't. The basis for frameworks like LangChain agents and the AutoGPT wave. Distinct from pure chain-of-thought (just thinking, no action). One of the most influential agent patterns. Related to tool use and agentic workflow.