LangGraph Research Agent

An autonomous multi-agent research system built with LangGraph that generates AI analyst personas, conducts web-based interviews, and produces a structured research report — all with human-in-the-loop feedback.

Graph Architecture

How It Works

The system is split into three cooperating graphs:

1. create_analysts — Analyst Generation

• Takes a research topic and a desired number of analysts (max_analysts)
• Uses GPT-4o to generate distinct AI analyst personas, each with a unique role, affiliation, and focus area
• Pauses via an interrupt to let you review and optionally refine the analysts before proceeding
• Loops back to regenerate if you provide feedback, or continues when you approve

2. question_answer_graph — Interview Subgraph

For each analyst, this subgraph runs an automated interview loop:

Node	What it does
ask_question	Analyst generates a targeted question based on their persona
search_web / search_web2	Two parallel Tavily web searches to gather source material
answer_question	Expert persona answers using retrieved context
save_interview	Serialises the full conversation transcript
write_section	Writes a structured markdown memo from the interview

The loop continues until max_num_turns expert responses are reached, then exits.

3. research_agent — Orchestrator

• Spawns a conduct_interview subgraph instance for each analyst in parallel using LangGraph's Send API
• Once all interviews are done, runs three nodes in parallel:
  • write_report — consolidates all analyst memos into a single narrative
  • write_introduction — writes a compelling intro with a title
  • write_conclusion — writes a crisp conclusion
• finalize_report assembles everything into the final markdown report

Project Structure

langraph-agent/
├── src/
│   ├── create_analysts.py       # Analyst generation graph
│   ├── answering_questions.py   # Interview subgraph
│   ├── agent.py                 # Main research orchestrator graph
│   └── utils/
│       ├── nodes.py             # All node functions
│       ├── edges.py             # Conditional edge functions
│       ├── states.py            # TypedDict state definitions
│       ├── objects.py           # Pydantic models (Analyst, Perspective, SearchQuery)
│       ├── prompts.py           # All prompt templates
│       ├── models.py            # LLM initialisation
│       └── tools.py             # Tool definitions
├── langgraph.json               # LangGraph server config
├── pyproject.toml
└── .env                         # API keys (not committed)

Setup

Requirements: Python 3.14+, uv

1. Clone the repo and install dependencies:

   uv sync

2. Create a .env file with your API keys:

   OPENAI_API_KEY=sk-...
   TAVILY_API_KEY=tvly-...
   LANGSMITH_API_KEY=lsv2-...   # optional, for tracing

3. Start the LangGraph dev server:

   uv run langgraph dev

4. Open LangGraph Studio at the URL printed in the terminal and run the research_agent graph with:

   {
     "topic": "The impact of AI on software engineering",
     "max_analysts": 3
   }

Key Technologies

• LangGraph — stateful multi-agent orchestration
• LangChain OpenAI — GPT-4o for generation
• Tavily — real-time web search
• LangGraph Studio — visual graph debugging