ReAct Pattern

ReAct (Reasoning + Acting) is a prompting framework for AI agents that interleaves chain-of-thought reasoning steps with discrete tool-use actions, allowing the model to observe each action's result and update its reasoning before proceeding.

ReAct is an agent architecture that alternates between two operation types: Reasoning (generating a natural-language thought about the current situation and what to do next) and Acting (executing a concrete operation such as a web search, database query, or code execution). The framework was introduced by Yao et al. in a 2022 paper, 'ReAct: Synergizing Reasoning and Acting in Language Models,' presented at ICLR 2023, and has since become one of the most widely adopted templates for constructing AI agents.

In a ReAct loop, the model produces a 'Thought' — its internal reasoning about the task state — followed by an 'Action' that invokes an external tool. The environment returns an 'Observation,' which is appended to the context window. The model then generates the next thought, informed by that observation, and the cycle repeats until the model judges the task complete. This interleaved structure allows the agent to correct course dynamically rather than committing upfront to a static plan that cannot incorporate new information.

The pattern addresses a core limitation of static chain-of-thought prompting: the model cannot acquire new external information mid-response. By coupling reasoning to real tool calls, ReAct enables tasks that require live data retrieval, interactive computation, or multi-step navigation through an external environment. The thought-action-observation trace is also human-readable, making it substantially easier to debug agent failures than opaque tool-calling pipelines where internal reasoning is hidden.

By 2026, ReAct-style loops underpin most production agent frameworks, including LangChain and LangGraph agents, Anthropic's Claude tool-use API, and OpenAI's Assistants API with code interpreter and file search. Extensions such as Reflexion add a self-critique and memory update step after task completion. Multi-agent variants allow one ReAct agent to spawn specialized sub-agents as its 'Actions.' The primary practical limitations are latency — each reasoning-acting cycle adds at least one full inference round trip — and context-window growth as traces accumulate over long tasks.