What is ReAct agent?

ReAct (Reasoning + Acting) is an agent pattern from a 2022 paper that combines chain-of-thought reasoning with tool use in a single loop. On each turn the model emits a thought (its reasoning about what to do next) and then an action (a tool call with arguments); the environment returns an observation (the tool's result), which is appended to the context, and the loop repeats. The reasoning traces help the model plan and recover from dead ends, while the actions ground that reasoning in real data instead of letting it hallucinate. ReAct is the conceptual ancestor of most modern tool-calling agents and of MCP-driven workflows: an MCP client exposes server tools to the model, the model reasons about which tool to call, calls it, reads the structured result, and continues. Compared with a fixed prompt chain, ReAct lets the model decide the path at runtime, which is powerful for open-ended tasks but harder to bound, so production deployments wrap it with step limits, guardrails, and human-in-the-loop approvals for risky actions. Frameworks like LangChain popularized ReAct as a default agent type before native tool-calling APIs made the pattern a first-class capability.