ReAct agent
An agent built on the ReAct pattern: an interleaved loop of reasoning (the model thinks out loud) and acting (the model calls a tool), repeated until the goal is met.
The ReAct pattern, introduced in the 2022 Yao et al. paper, is the canonical agent design. Each turn the LLM emits two things: a reasoning step ("I should look up the user's account") and an action ("call get_account(user_id=42)"). The action runs; the observation feeds back into the next reasoning step. The loop continues until the model emits a final answer.
Almost every production agent in 2026 is a ReAct agent under the hood. The variations — extended thinking, planning steps before action, parallel tool calls, reflection passes — are all evolutions of the same core skeleton. What differentiates good agents is the reasoning quality at each step, not the loop itself.
For builders, ReAct is the right starting point for new agents. It is dead-simple to implement (a system prompt that demonstrates the format), it works with any tool-calling-capable model, and it gives you observable traces of the agent's reasoning. Add complexity (planning, multi-agent) only when ReAct alone hits a ceiling.
Frequently asked
Is every agent a ReAct agent?+
In practice, almost. Most modern agents implement the reason-act-observe loop even if they do not use the ReAct label. Variations add planning, reflection, or multi-agent layers on top of the same skeleton.
How is ReAct different from chain of thought?+
CoT is reasoning-only; the model thinks out loud and produces an answer. ReAct interleaves reasoning with tool calls — think, act, observe, think again. ReAct extends CoT into the agent setting.