Two-Step Reasoning & Action Agent

Overview

This project demonstrates a simple two-step conversational agent built using LangChain and LangGraph. It separates the reasoning process into two distinct stages:

1. Reasoning Stage: Generates an intermediate "thought" by prompting the LLM to think step by step about the user input.
2. Action Stage: Constructs a concrete action plan based on the original input and the generated "thought".

A state machine (StateGraph) orchestrates these stages in sequence and produces both the reasoning trace and final plan.

---

Features

• Modular agent design with clear separation of reasoning and planning.
• Uses Pydantic models (AgentState) to define and validate agent state.
• Easy-to-extend workflow graph via StateGraph.
• Visualization of the workflow as a Mermaid diagram.

---

Requirements

• Python 3.8+
• langchain
• langgraph
• pydantic
• openai (for the ChatOpenAI integration)

Install dependencies via pip:

pip install langchain langgraph pydantic openai

---

Usage

1. Define the Agent State

   • AgentState captures input_text, thought, and output_action.

2. Implement Nodes

   • reasoning_agent: Prompts the LLM to think step by step about the input.
   • action_agent: Builds an action plan using both input and intermediate thought.

3. Build the Workflow

   • Instantiate StateGraph, add nodes, set entry point, and define transitions.
   • Compile the graph to obtain an executable workflow.

4. Invoke the Workflow

   • Create an initial AgentState with the user request.
   • Call graph.invoke(initial_state) to execute reasoning and planning.

5. Visualize the Graph

   • Generate a Mermaid PNG diagram via graph.get_graph(xray=1).draw_mermaid_png().

Example:

# Prepare initial state
initial_state = AgentState(
    input_text="Build a 3-month plan to study the basics of English language",
    thought="",
    output_action=""
)

# Run the workflow
result = graph.invoke(initial_state)
print(result.thought)
print(result.output_action)

# Save workflow diagram
png = graph.get_graph(xray=1).draw_mermaid_png()
open("workflow.png", "wb").write(png)

---

Architecture

flowchart TD
  A[User Input] --> B[Reasoning Agent]
  B --> C[Action Agent]
  C --> D[Final Plan]

Loading

• StateGraph manages transitions between nodes.
• ChatOpenAI provides the LLM backend for both stages.

---

Extending the Agent

• Add more nodes for validation, execution, or feedback loops.
• Customize prompts or integrate additional tools.
• Adjust state schema to capture more context as needed.

---

License

This project is released under the MIT License.