ReAct Prompting

ReAct (Reasoning + Acting) is a prompting pattern where the AI alternates between thinking about what to do and actually doing it. Instead of reasoning in isolation and then acting, or acting without reasoning, the AI interleaves thought and action in a loop.

The pattern

1. Thought: "I need to find the current price of X. Let me search for it."
2. Action: Search for "current price of X"
3. Observation: "The search returned £42.50 as of today."
4. Thought: "Now I need to compare this to last month's price. Let me search for that."
5. Action: Search for "price of X last month"
6. Observation: "Last month it was £38.00."
7. Thought: "That is a 12% increase. I can now answer the user's question."
8. Final answer: "X has increased 12% from £38.00 to £42.50 over the past month."

Why it matters

ReAct is the foundation of how modern AI agents work. The agent loop (plan → act → observe → decide) described in agentic systems is essentially ReAct at scale. When Claude Code reads a file, runs a test, checks the result, and decides what to do next — that is ReAct in action.

When to use it

ReAct is most valuable when the AI needs external information to complete a task. Pure reasoning tasks (logic puzzles, creative writing) do not benefit. Tasks requiring real-world data, multi-step research, or verification of claims benefit significantly.

As a prompting technique

You can trigger ReAct-like behaviour even without tool access by prompting: "Think step by step. For each step, state what you would do, what information you would need, and what you would expect to find. Then proceed to the next step based on your reasoning."