ReAct Pattern Revealed

ReAct (Reasoning + Acting) is the dominant pattern for LLM agents, combining chain-of-thought reasoning with tool-use actions.

The Loop

Each agent step:

1. Thought: Reasoning about current state and next action
2. Action: Tool call (search, code execution, API) or final answer
3. Observation: Environment result from the action
4. Repeat until task completion

Implementation

LangChain and LangGraph provide ReAct implementations. Tools are defined as functions with name, description, parameters, and executor. The agent receives tool descriptions and selects which to call based on reasoning.

Best Practices

Limit iterations to 10-15 to prevent infinite loops. Handle tool failures gracefully with retry or fallback. All steps should be logged for debugging. The final answer should cite observations as evidence.