ReAct: 도구를 활용한 추론과 실행

Chain-of-thought helps a model reason, but reasoning alone cannot fetch live data or run code. The ReAct pattern combines Reasoning with Acting so the model can use tools mid-task.

ReAct alternates three roles in a loop:

• Thought: reason about what to do next.
• Action: call a tool with an input.
• Observation: read the tool's result.

Repeat until the answer is ready.

A typical trace looks like this. The model emits a thought, an action, then waits for an observation injected back into the prompt.

You describe the available tools in the prompt: their names, what they do, and the input format. The model picks among them.

In practice your code stops generation when it sees an Action, executes the real tool, then appends the Observation and resumes the model. The model never invents tool results itself.

Because the model grounds each step in real observations, ReAct cuts hallucination on factual or computational tasks. The reasoning explains why each tool was called, aiding debugging.

ReAct shines on tasks needing several lookups: find a company, then its founder, then that person's birth year. Each step's observation feeds the next thought.

Always cap the number of iterations. Without a limit, a confused model can loop forever calling tools. Add a max-step budget and a fallback answer.

Tools fail: timeouts, bad inputs, empty results. Feed the error back as an observation so the model can retry differently rather than crashing the loop.

• CoT: reasons internally, no external data.
• ReAct: reasons AND acts on the world via tools.

Use ReAct whenever the answer depends on information the model cannot already know.

ReAct is the backbone of modern LLM agents. Frameworks like LangChain implement this thought-action-observation loop under the hood to build autonomous tool-using assistants.

ReAct interleaves Thought, Action, and Observation so the model reasons and uses real tools. It grounds answers, reduces hallucination, needs a step limit and error handling, and underpins modern LLM agents.