GraphNews: A Multi-Agent Self-Correcting Editorial Team

GraphNews is a sophisticated multi-agent system built with LangGraph that automates the end-to-end process of news research, curation, and newsletter drafting. Unlike standard linear AI chains, this project utilizes a Cyclic Graph architecture to implement a "Self-Correction" loop, ensuring high-quality, factual output through autonomous peer review.

Workflow Architecture

The system treats the editorial process as a state machine. Data flows through specialized nodes, maintaining a shared state to ensure context is preserved across the entire lifecycle:

1. RESEARCHER: Scours the web using Tavily for the top 10 .
2. CURATOR: Filters research results to select the top 5 most relevant sources.
3. WRITER: Generates a structured Markdown draft based on curated items.
4. CRITIC: Evaluates the draft. If the score is < 8/10, the draft is sent back to the WRITER with specific feedback for revision.
5. FINALIZE: Once the quality threshold is met, or we reach the 3rd draft (to avoid infinite loops), the final newsletter is saved to disk.

graph TD
    Start((Start)) --> Researcher[Researcher Node]
    Researcher --> Curator[Curator Node]
    Curator --> Writer[Writer Node]
    Writer --> Critic[Critic Node]
    
    Critic --> Decision{Score >= 8?}
    
    Decision -- "No (Score < 8 and draft number < 3)" --> Writer
    Decision -- "Yes (Approved)" --> End((Save Markdown))

    subgraph "The Agentic Loop"
    Writer
    Critic
    Decision
    end

    style Researcher fill:#f9f,stroke:#333,stroke-width:2px
    style Curator fill:#bbf,stroke:#333,stroke-width:2px
    style Writer fill:#dfd,stroke:#333,stroke-width:2px
    style Critic fill:#fdd,stroke:#333,stroke-width:2px
    style Decision fill:#fff,stroke:#333,stroke-width:2px

Loading

Tech Stack

• Orchestration: LangGraph (Stateful Multi-Agent Workflows)
• LLM: DeepSeek-V3.2
• Search Engine: Tavily AI (Search API for LLMs)
• Language: Python

Quick Start

1. Clone & Install

git clone https://github.com/simonsl07/GraphNews.git
cd GraphNews
pip install -r requirements.txt

2. Create a .env file following the .env.example file

DEEPSEEK_API_KEY=your_key_here
TAVILY_API_KEY=your_key_here

3. Run the system

python main.py

Technical Challenges & Solutions

• The "Hallucination" Brake: Implemented a max revision counter within the Critic node logic to prevent infinite loops and runaway API costs.
• Deterministic State Control: Used LangGraph's TypedDict state to maintain a single source of truth, preventing context drift.

Roadmap

• Add CLI arguments to control news topic and source count.
• Integrate LangSmith for full-trace agent debugging and cost monitoring.
• Human-in-the-Loop: Add a LangGraph Checkpointer to pause for human approval before final output.