ReAct (Reasoning + Acting)

ReAct is the pattern that made AI agents actually work. The name combines "Reason" and "Act," and the idea is simple: instead of asking the model to plan everything up front and then execute, you have it alternate between thinking and doing in a tight loop. The model writes a thought ("I need to find the user's recent order history"), takes an action ("call the get_orders tool with user_id=123"), receives an observation (the tool's output), then writes another thought, another action, and so on until it decides it's done. This loop is the foundation of nearly every modern agent framework — LangChain agents, AutoGPT, OpenAI's function calling loops, Claude's tool use, MCP-based agents. The loop structure matters because real-world tasks require adapting to what you find. A research agent doesn't know in advance which links it'll need to follow. A coding agent doesn't know which tests will fail. ReAct lets the model decide its next move based on what just happened, instead of committing to a rigid plan. The original 2022 paper showed ReAct beat both pure chain-of-thought (which can't act) and pure tool-using agents (which can't reason about results). By 2026, ReAct is mostly invisible — frameworks and APIs implement it under the hood, often with extensions like reflection (the model critiques its own actions), memory (persisting context across loops), and parallel tool calls (multiple actions per step). But every agent you talk to is still, at its core, doing ReAct.