Multi-Agentic Deep Research System

A sophisticated, asynchronous multi-agent AI system built with LangGraph, LangChain, and Tavily Search to autonomously scope, research, and synthesize comprehensive reports on any topic.

🚀 Overview

This project uses a pipeline of specialized AI agents working together to conduct deep research. Instead of relying on a single LLM prompt, the workload is distributed across four distinct agents that handle user scoping, task delegation, parallel web research, and final report synthesis.

🤖 The Agent Architecture

The system operates in three distinct phases, coordinated sequentially by a main entry script (chat.py).

1. Scoping Agent (scoping/)

Purpose: To act as the conversational frontend and define the exact parameters of the research.

• How it works: It engages in an interactive chat with the user. If a request is vague, it asks clarifying questions using conversation memory (checkpointer). Once it has enough context (or after 5 clarification attempts), it translates the chat history into a highly specific Research Brief.
• Handoff: The Research Brief is passed downstream as the master instruction manual for the rest of the pipeline.

2. Supervisor Agent (research_supervisor/)

Purpose: To orchestrate the research process and distribute workloads.

• How it works: The Supervisor reads the Research Brief and breaks it down into multiple sub-topics. It then asynchronously spawns parallel Researcher Agents to investigate each sub-topic simultaneously, drastically reducing research time.

3. Researcher Agent (research/)

Purpose: To conduct deep, iterative web research.

• How it works: Each Researcher is given a specific sub-topic by the Supervisor. It operates in a tool-calling loop using the Tavily Search API. After every search, it uses an internal think_tool to reflect on the findings, identify missing information, and formulate its next query. Once satisfied (or hitting a hard query limit), it cleans and returns its findings to the Supervisor.

4. Reporter Agent (reporter/)

Purpose: To synthesize a polished final deliverable.

• How it works: Once all parallel Researchers return their findings to the Supervisor, the aggregated raw notes are passed to the Reporter Agent. It structures the data, removes redundancies, and generates a comprehensive, cleanly formatted Markdown report complete with inline citations and a sources list.

🛠️ Setup & Execution

Prerequisites

• Python 3.9+
• A .env file in the root directory containing your API keys:

  GROQ_API_KEY=your_groq_key
  GEMINI_API_KEY=your_gemini_key
  TAVILY_API_KEY=your_tavily_key
  
  # Optional: For LangSmith tracing
  LANGCHAIN_TRACING_V2=true
  LANGCHAIN_API_KEY=your_langchain_key
  LANGCHAIN_PROJECT="Research Agent"

Running the System

To run the full end-to-end pipeline:

python chat.py

Standalone Testing

You can also test the Scoping or Research agents independently by running their standalone scripts:

python scoping/chat.py
python research/chat.py

📊 Observability

The entire system is instrumented for LangSmith. Because each node and sub-graph is explicitly named (e.g., Scoping_Agent, Supervisor_Agent, Parallel_Worker_xyz), you can monitor the exact execution traces, tool calls, and latency of parallel agents directly in your LangSmith dashboard.