Flashcard: what is the ReAct pattern for agents, in one sentence?

**ReAct** stands for Reason and Act. It is a pattern where an LLM agent takes turns between thinking and doing. First the model writes a Thought, a short sentence explaining what it is reasoning about. Then it emits an Action, which is usually a tool call. The runtime executes that tool and writes the result back into the conversation as an Observation. The model then writes the next Thought, and the cycle repeats. The loop ends when the model decides it has enough information and emits a Final Answer instead of another Action. Why this matters. A plain LLM call is a single forward pass. It generates a long response in one go, with nothing from the outside world entering along the way. ReAct breaks that pass into many smaller passes, with real data injected between each one. Each Thought is conditioned on the latest Observation, so the model is reasoning from facts it just saw, not from facts it imagined. This is the pattern almost every modern agent framework uses under the hood: LangGraph, CrewAI, AutoGen, and the provider-native tool-use loops in Claude and GPT-5.5.

**ReAct** is the pattern of interleaving Reasoning and Acting in an LLM agent. The model emits a Thought, then an Action (typically a tool call), then receives an Observation, which is the real result of that Action. The cycle repeats: Thought, Action, Observation, Thought, Action, Observation, until the model emits a Final Answer. The interleaving is the entire point. Because each Observation is written into the context before the next Thought is sampled, the next reasoning step is conditioned on real data instead of the model's own continuation. A single forward pass has no such checkpoint; the model generates the whole trace from its weights and never pauses to verify anything against the outside world. The pattern was introduced by Yao et al. in 2022 (arXiv 2210.03629) and became the dominant substrate for tool-using agents because the closed loop bounds hallucination drift on long tasks. A wrong assumption gets corrected at the next Observation rather than carried to the end. In an interview, the one-sentence answer is: ReAct alternates Thought, Action, Observation in a loop, so each reasoning step stands on a real tool result rather than the model's own guess. That interleaving is what makes it an agent loop instead of a single forward pass.