Skip to main content

Assignment: Multi-Expert ReAct Research Agent

Assignment Metadata​

FieldDescription
Assignment NameMulti-Expert ReAct Research Agent
CourseLangGraph and Agentic AI
Project Namereact-research-agent
Estimated Time120 minutes
FrameworkPython 3.10+, LangGraph, LangChain, OpenAI API

Learning Objectives​

By completing this assignment, you will be able to:

  • Understand the ReAct pattern (Reasoning + Acting) for agent design
  • Implement LLM-based expert tools instead of simple web search
  • Use LangGraph's prebuilt ToolNode for tool execution
  • Apply iteration control to prevent infinite loops
  • Design coordinator prompts that enable structured reasoning

Problem Description​

You are building a research agent that can consult multiple domain experts to answer complex questions. Instead of using web search, the agent uses specialized LLM experts:

  • AI Research Expert: Specializes in ML papers, architectures, trends
  • Financial Analyst: Specializes in stocks, markets, valuations

The coordinator LLM uses ReAct pattern to decide which expert(s) to consult and synthesize their responses.

Technical Requirements​

Environment Setup​

  • Python 3.10 or higher
  • Required packages:
    • langgraph >= 0.2.0
    • langchain >= 0.1.0
    • langchain-openai >= 0.1.0

API Requirements​

  • OpenAI API key configured as environment variable

Tasks​

Task 1: Expert Tool Definition (25 points)​

  1. Create AI Research Expert tool that:

    • Uses a specialized system prompt for ML/AI topics
    • Invokes an LLM with research-focused instructions
    • Returns structured analysis with paper-style references

  2. Create Financial Analyst tool that:

    • Uses a specialized system prompt for market analysis
    • Provides data-driven insights with market context
    • Returns actionable financial information

Task 2: Coordinator Implementation (30 points)​

  1. Define ResearchState with:

    • messages: For conversation history
    • max_iterations: Limit for ReAct loops (default: 5)
    • current_iteration: Tracking current loop count

  2. Implement coordinator_node that:

    • Binds expert tools to the coordinator LLM
    • Includes ReAct reasoning in the prompt
    • Increments iteration counter on each call

  3. Create routing function (should_continue) that:

    • Checks for tool_calls in the last message
    • Enforces max_iterations limit
    • Routes to "tools" or "end" appropriately

Task 3: Graph Construction (25 points)​

  1. Use prebuilt ToolNode for tool execution (NOT custom implementation)

  2. Build the graph with:

    • Coordinator node as entry point
    • ToolNode for expert execution
    • Conditional edges for ReAct loop
    • Edge from tools back to coordinator

  3. Configure checkpointer for state persistence

Task 4: Testing Multi-Step Reasoning (20 points)​

  1. Test with complex queries requiring multiple experts:

    • "How are AI companies valued in the current market?"
    • "Compare NVIDIA vs AMD for AI workloads - technical and market position"

  2. Verify ReAct flow:

    • Iteration 1: First expert consultation
    • Iteration 2+: Additional consultations or synthesis
    • Final: Comprehensive answer combining perspectives

  3. Document the reasoning chain showing Think → Act → Observe steps

Submission Requirements​

Required Deliverables​

  • Source code in Python script or Jupyter notebook
  • README.md with setup instructions
  • Test output showing multi-step ReAct reasoning
  • Graph visualization

Submission Checklist​

  • Both expert tools implemented with specialized prompts
  • ToolNode from langgraph.prebuilt is used (not custom)
  • Iteration control prevents infinite loops
  • Coordinator successfully synthesizes multi-expert input
  • Code runs without errors

Evaluation Criteria​

CriteriaPoints
Expert tool implementation25
Coordinator with ReAct prompt30
Graph construction with ToolNode25
Multi-step reasoning tests15
Code quality and documentation5
Total100

Hints​

tip
  • Use llm.bind_tools([tool1, tool2]) to enable tool calling for the coordinator
  • The prebuilt ToolNode(tools) handles parsing, execution, and ToolMessage creation
  • Include few-shot examples in coordinator prompt to guide tool selection
  • Set temperature=0.3 for expert LLMs for consistent responses

References​