Skip to main content

Lesson 2 · 11 min

ReAct — Reason + Act + Observe

The pattern under almost every modern agent. Surprisingly simple, surprisingly effective.

The original idea

ReAct (Reason + Act, 2022) is the pattern: at each turn, the model writes:

  1. A Thought: what it knows, what it needs.
  2. An Action: which tool to call, with what arguments.
  3. Receives an Observation: the tool's output.

Then it loops. Eventually it writes a final answer instead of an action.

This is how Claude, GPT, Gemini, Grok all run agents under the hood. The frontier model providers expose it as tool use in the API.