Multi-Agent Research Assistant

A LangGraph-based research pipeline that takes any question, breaks it into focused sub-tasks, runs parallel web searches, summarizes findings, fact-checks claims, and produces a structured Markdown report.

Built with LangGraph, the Anthropic API, and Python.

---

Architecture

research_question
       |
  orchestrator        Decomposes the question into 3 focused sub-tasks
       |
  web_search          Runs a live web search for each sub-task
       |
  summarizer          Condenses raw results into summaries + key points
       |
  fact_checker        Cross-checks all summaries for consistency
       |
  compiler            Writes the final structured Markdown report
       |
  output/*.md         Report saved to disk

Each stage is a LangGraph node that reads from and writes to a shared ResearchState TypedDict. If any stage fails, the graph routes directly to the compiler which produces an error report rather than crashing.

---

Project Structure

research_assistant/
│
├── main.py                  # CLI entry point
├── graph.py                 # LangGraph StateGraph definition
├── state.py                 # Shared TypedDict state schema
│
├── agents/
│   ├── __init__.py
│   ├── orchestrator.py      # Decomposes question into 3 sub-tasks
│   ├── web_search.py        # Calls Anthropic API with web_search tool
│   ├── summarizer.py        # Condenses raw results per sub-task
│   ├── fact_checker.py      # Cross-checks claims across all summaries
│   └── compiler.py          # Assembles the final Markdown report
│
├── tools/
│   ├── __init__.py
│   └── search_tool.py       # Web search tool config and helpers
│
├── prompts/
│   ├── orchestrator.txt
│   ├── summarizer.txt
│   ├── fact_checker.txt
│   └── compiler.txt
│
├── output/                  # Generated reports saved here
├── requirements.txt
├── .env.example
└── README.md

---

Setup

1. Clone and enter the project directory:

git clone https://github.com/ashbix23/Multi-Agent_Research_Assistant
cd research_assistant

2. Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate        # Mac/Linux
venv\Scripts\activate           # Windows

3. Install dependencies:

pip install -r requirements.txt

4. Configure environment variables:

cp .env.example .env

Open .env and add your Anthropic API key:

ANTHROPIC_API_KEY=sk-ant-your-key-here
RESEARCH_MODEL=claude-sonnet-4-20250514
MAX_SEARCH_RESULTS=5
OUTPUT_DIR=output

Get your API key from console.anthropic.com.

---

Usage

Pass a question directly as a CLI argument:

python main.py "What is the current state of nuclear fusion energy?"

Or run interactively:

python main.py
# Research question: <you type here>

Import and run programmatically:

from main import run

state = run("What are the geopolitical implications of Arctic ice melt?")
print(state["final_report"])

---

Output

Each run produces a Markdown report saved to the output/ directory:

output/
└── 20250406_143022_Nuclear_Fusion_Energy.md

The report follows this structure:

# Title

## Executive Summary
## Findings
   ### Sub-task 1 focus
   ### Sub-task 2 focus
   ### Sub-task 3 focus
## Key Takeaways
## Reliability Assessment
## Sources

The terminal also displays:

• Live progress from each agent as it runs
• Word count and file path on completion
• Confidence verdict (HIGH / MEDIUM / LOW) from the fact-checker
• Any errors encountered during the pipeline

---

Configuration

Variable	Default	Description
ANTHROPIC_API_KEY	—	Required. Your Anthropic API key
RESEARCH_MODEL	claude-sonnet-4-20250514	Model used by all agents
MAX_SEARCH_RESULTS	5	Max web search calls per sub-task
OUTPUT_DIR	output	Directory where reports are saved

---

How Each Agent Works

Orchestrator

Sends the raw research question to Claude and receives back a JSON object containing a title and 3 sub-tasks. Each sub-task has a specific search query and a focus description. This decomposition drives the rest of the pipeline.

Web Search Agent

Calls the Anthropic API with the built-in web_search_20250305 tool enabled. For each sub-task, it runs a targeted search and collects the raw response text and any source URLs. Searches run sequentially to avoid rate limits.

Summarizer Agent

Takes each raw search result and sends it to Claude with instructions to return a structured JSON summary (150–250 words) plus 3–5 key points. Preserves specific facts, figures, and dates rather than paraphrasing them away.

Fact Checker Agent

Receives all 3 summaries at once and cross-checks them for consistency. Returns a verdict of high, medium, or low confidence along with lists of verified and uncertain claims. This is the only agent that sees all summaries simultaneously.

Compiler Agent

Receives all summaries, the fact-check result, and all collected sources, then writes the full Markdown report. Uses max_tokens=4096 since it is producing a full document rather than compact JSON. Saves the report to disk with a timestamped filename.

---

---

Requirements

• Python 3.11+
• An Anthropic API key with access to Claude Sonnet 4
• Internet access for the web search tool