The ReAct Pattern: How AI Agents Reason and Act in Alternating Steps

What Is the ReAct Pattern?

ReAct stands for Reasoning + Acting. Introduced in a 2022 research paper by Yao et al., it is the foundational pattern behind nearly every modern AI agent. The core idea is simple: instead of having an LLM generate a final answer in one shot, you let it alternate between thinking (reasoning) and doing (acting), with observations from the environment feeding back into the next reasoning step.

This alternation is what gives agents their power. A single LLM call might hallucinate an answer. A ReAct loop grounds the LLM's reasoning in real data from real tools.

The ReAct Cycle

Each iteration of a ReAct loop has three phases:

flowchart TD
    USER(["User message"])
    LLM["LLM call<br/>with tools schema"]
    DECIDE{"Model wants<br/>to call a tool?"}
    EXEC["Execute tool<br/>sandboxed runtime"]
    RESULT["Append tool_result<br/>to messages"]
    GUARD{"Output passes<br/>guardrails?"}
    DONE(["Final reply"])
    BLOCK(["Refuse and log"])
    USER --> LLM --> DECIDE
    DECIDE -->|Yes| EXEC --> RESULT --> LLM
    DECIDE -->|No| GUARD
    GUARD -->|Yes| DONE
    GUARD -->|No| BLOCK
    style LLM fill:#4f46e5,stroke:#4338ca,color:#fff
    style EXEC fill:#ede9fe,stroke:#7c3aed,color:#1e1b4b
    style GUARD fill:#f59e0b,stroke:#d97706,color:#1f2937
    style DONE fill:#059669,stroke:#047857,color:#fff
    style BLOCK fill:#dc2626,stroke:#b91c1c,color:#fff

1. Thought — The LLM reasons about the current state and decides what to do next
2. Action — The LLM calls a tool or function
3. Observation — The tool returns a result, which feeds into the next thought

User: "What is the current stock price of AAPL and is it above its 50-day average?"

Thought 1: I need to look up the current AAPL stock price first.
Action 1: get_stock_price(symbol="AAPL")
Observation 1: {"price": 187.42, "currency": "USD", "timestamp": "2026-03-17T14:30:00Z"}

Thought 2: The current price is $187.42. Now I need the 50-day moving average.
Action 2: get_moving_average(symbol="AAPL", days=50)
Observation 2: {"average": 182.15, "period_days": 50}

Thought 3: AAPL is at $187.42, above its 50-day average of $182.15. I can answer now.
Action 3: respond("AAPL is currently at $187.42, which is above its 50-day
           moving average of $182.15 by approximately $5.27 or 2.9%.")

Notice how each thought is informed by the previous observation. The agent does not guess — it retrieves real data and then reasons about it.

Why ReAct Beats Chain-of-Thought Alone

Chain-of-Thought (CoT) prompting asks the LLM to think step by step, but all reasoning happens internally without access to external information. This means CoT is limited by what the model already knows (its training data), and it can confidently produce wrong answers.

ReAct fixes this by interleaving actions between thoughts. The LLM can verify its assumptions, fetch current data, and correct course based on real observations.

Approach	Grounded in real data?	Multi-step?	Can use tools?
Direct prompting	No	No	No
Chain-of-Thought	No	Yes (internal)	No
ReAct	Yes	Yes	Yes

Implementing ReAct in Python

Here is a working ReAct agent using the OpenAI API with function calling:

import json
import openai

client = openai.OpenAI()

# Define the tools the agent can use
tools = [
    {
        "type": "function",
        "function": {
            "name": "search_web",
            "description": "Search the web for current information",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string", "description": "Search query"}
                },
                "required": ["query"],
            },
        },
    },
    {
        "type": "function",
        "function": {
            "name": "calculate",
            "description": "Evaluate a mathematical expression",
            "parameters": {
                "type": "object",
                "properties": {
                    "expression": {"type": "string", "description": "Math expression"}
                },
                "required": ["expression"],
            },
        },
    },
]

def execute_tool(name: str, arguments: dict) -> str:
    """Execute a tool and return the result as a string."""
    if name == "search_web":
        # In production, call a real search API
        return json.dumps({"results": [f"Result for: {arguments['query']}"]})
    elif name == "calculate":
        try:
            result = eval(arguments["expression"])  # Use a safe evaluator in production
            return json.dumps({"result": result})
        except Exception as e:
            return json.dumps({"error": str(e)})
    return json.dumps({"error": f"Unknown tool: {name}"})

def react_agent(user_input: str, max_iterations: int = 10) -> str:
    messages = [
        {"role": "system", "content": "You are a helpful agent. Use tools to find "
         "accurate information before answering. Think step by step."},
        {"role": "user", "content": user_input},
    ]

    for i in range(max_iterations):
        response = client.chat.completions.create(
            model="gpt-4o",
            messages=messages,
            tools=tools,
        )

        choice = response.choices[0]
        messages.append(choice.message)

        # If no tool calls, the agent is done reasoning
        if not choice.message.tool_calls:
            return choice.message.content

        # Execute each tool call (Action) and add results (Observation)
        for tool_call in choice.message.tool_calls:
            args = json.loads(tool_call.function.arguments)
            result = execute_tool(tool_call.function.name, args)

            messages.append({
                "role": "tool",
                "tool_call_id": tool_call.id,
                "content": result,  # This is the Observation
            })

    return "Agent reached maximum iterations without a final answer."

# Run the agent
answer = react_agent("What is 15% of the population of France?")
print(answer)

The key insight is on line where we check if not choice.message.tool_calls. When the LLM decides it has enough information, it responds with text instead of a tool call — that is how ReAct agents naturally terminate.

When ReAct Falls Short

ReAct is powerful but not perfect. It struggles with tasks that require long-horizon planning (more than 8-10 steps), because LLMs can lose track of their overall goal as context grows. For those cases, you layer planning patterns on top of ReAct — decompose the task into subtasks first, then use ReAct for each subtask.

ReAct also depends entirely on tool quality. If your tools return ambiguous or inconsistent results, the agent's reasoning degrades. Investing in clear tool descriptions and reliable tool implementations is just as important as prompt engineering.

FAQ

What is the difference between ReAct and function calling?

Function calling is the mechanism (how the LLM requests a tool execution). ReAct is the pattern (alternating reasoning and acting in a loop). ReAct uses function calling as its action mechanism, but adds the structured loop of thought-action-observation that continues until the task is complete.

How many iterations should a ReAct loop allow?

For most tasks, 5 to 15 iterations is sufficient. Set a maximum to prevent runaway loops and cost explosions. If your agent consistently hits the limit, the task is likely too complex for a single ReAct loop — consider decomposing it into subtasks instead.

Can ReAct work with open-source models?

Yes. Any LLM that supports function calling or structured output can run a ReAct loop. Models like Llama 3, Mistral, and Qwen all support tool use. The quality of reasoning depends on model capability, but the pattern itself is model-agnostic.

---

#ReAct #AIAgents #Reasoning #ToolUse #Python #AgenticAI #LearnAI #AIEngineering