ReAct agents alternate between thinking and acting. The LLM first reasons (Thought), then decides what tool to call (Action), observes the result (Observation), and repeats. This pattern outperforms both "pure reasoning" and "pure acting" on complex tasks.
Thought-Action-Observation Loop
Thought: "I need to find the population of France in 2024."
Action: search_tool("France population 2024") Observation: "67.8 million (as of 2024)" Thought: "Now I have the answer. I should provide it with context." Action: provide_answer("67.8 million")
Implementing ReAct
from langchain_openai import ChatOpenAI
from langchain.agents import initialize_agent, Tool, AgentType
tools = [ Tool( name="search", func=search_function, description="Search the internet" ), Tool( name="calculate", func=calculator_function, description="Perform arithmetic" ) ]
agent = initialize_agent( tools=tools, llm=ChatOpenAI(model="gpt-4"), agent=AgentType.REACT_DOCSTRING, verbose=True )
result = agent.run("What is the population of France divided by the population of Germany?")
The agent will reason, search for both populations, calculate the ratio, and return the answer.
Conclusion
ReAct agents effectively combine reasoning and tool use. Interleaving thought and action prevents both hallucination (from pure reasoning) and missed reasoning steps (from pure acting). Understanding the ReAct pattern is essential for building effective agents. Next: we'll explore multi-agent orchestration—how to coordinate multiple specialized agents.