What it does GPT Newspaper is a Python app that runs a pipeline of specialized agents to build a personalized newspaper from scratch. You pick topics and sources; the system searches the web, selects stories, writes articles, critiques them, lays them out, and publishes to a browser view. It’s essentially a small editorial room where every employee is an LLM.

The interesting bit The architecture uses LangChain’s LangGraph to wire seven single-purpose agents into a workflow — including a Critique Agent that loops feedback to the Writer until the article passes muster. That’s a rare touch of actual process control in a space that usually just chains prompts and hopes.

Key highlights

• Seven-agent pipeline: Search → Curator → Writer → Critique → Designer → Editor → Publisher
• Built on LangGraph, not just a linear prompt chain
• Web search via Tavily, generation via OpenAI
• Simple Flask(ish) local server at localhost:5000
• Demo video shows a rendered multi-article newspaper page

Caveats

• Requires two paid API keys (Tavily + OpenAI) to run at all
• README claims “six specialized sub-agents” but lists seven; the count mismatch is right there in the overview
• Explicitly labeled “experimental” and “not a replacement for professional news outlets” — so hallucinations and source drift are expected hazards, not bugs

Verdict Worth a spin if you’re building multi-agent workflows and want a concrete, visual example of role-based LLM collaboration. Skip it if you need production-grade news aggregation or are allergic to API metered billing.