ReAct: Synergizing Reasoning and Acting in Language Models

Summary

Shows that interleaving reasoning traces ("thought") with actions (tool calls) and observations lets a language model solve tasks more reliably than reasoning alone or acting alone.

Key points

• Loop: think, act, observe, repeat until done.
• Reasoning helps plan and recover from errors; actions ground the model in real data.
• Reduces hallucination on knowledge tasks compared with chain-of-thought alone.

Why it matters

The think-act-observe loop is the basic shape of almost every modern agent, including tool-use loops in vendor APIs.