How does ReAct (Reasoning and Acting) prompting work?

Short answer: ReAct (Reasoning + Acting) is a prompting technique that combines Chain-of-Thought reasoning with external action/tool execution in an interleaved loop (Thought → Action → Observation → Thought), allowing LLMs to solve complex tasks dynamically.

Detail

Standard Chain-of-Thought (CoT) prompting encourages the model to generate internal reasoning steps before outputting a final answer. However, CoT is limited by the model's static parameter memory and cannot inspect real-world databases, call APIs, or handle dynamic environment changes.

ReAct bridges reasoning and execution by structuring prompt generation into discrete steps:

1. Thought: The model analyzes the current context, sub-goal, and what information is missing.
2. Action: The model emits a structured command or tool call (e.g., Search[query], ExecuteSQL[query], Calculator[expression]).
3. Observation: The external system runs the tool and appends the result back into the conversation context as an Observation.
4. Repeat: The model reads the new observation and generates the next Thought, continuing until it reaches a final Thought and outputs Finish[answer].

User Query
    │
    ▼
┌───────┐
│Thought│ ──► "I need to check stock price of Apple."
└───────┘
    │
    ▼
┌───────┐
│Action │ ──► GetStockPrice("AAPL")
└───────┘
    │
    ▼
┌───────────┐
│Observation│ ──► "$224.50" (Result returned by API)
└───────────┘
    │
    ▼
┌───────┐
│Thought│ ──► "Now I have the price. I can return the final answer."
└───────┘
    │
    ▼
  Finish

Why ReAct is central to AI Engineering

• Explainability: Human developers can inspect the model's reasoning trajectory step-by-step.
• Error Recovery: If an action returns an error or empty result, the subsequent Thought step allows the LLM to modify its search query or try an alternative tool.
• Groundedness: Reduces hallucinations by grounding answers in retrieved observations rather than parametric memory.

Example

Prompt template structure for a ReAct agent:

You operate in a loop of Thought, Action, Observation, Thought.
Available Tools:
- Search[query]: Searches the internet for information.
- Calculator[expression]: Evaluates a mathematical expression.

Format your response strictly as follows:
Thought: <reasoning about what to do>
Action: <ToolName[argument]>

When you have the final answer, format it as:
Thought: I know the final answer.
Final Answer: <your response>

User Question: What is the age of the current President of France multiplied by 2?

Thought: First I need to find out who the current President of France is and their age.
Action: Search[current President of France age]
Observation: Emmanuel Macron, born 21 December 1977 (age 48 as of August 2026).

Thought: I found Emmanuel Macron's age (48). Now I need to multiply it by 2.
Action: Calculator[48 * 2]
Observation: 96

Thought: I have calculated the final result.
Final Answer: The age of the current President of France multiplied by 2 is 96.

Interview tips

• Contrast ReAct with pure zero-shot tool calling (Function Calling API): ReAct explicitly surfaces reasoning tokens prior to tool execution, improving tool argument selection accuracy.
• Discuss common failure modes: infinite loops (model repeatedly trying broken tools), context window overflow from long observation outputs, and high API costs.

---

⬅ Back to Prompt Engineering · All topics