Newsloop

interface
workflow
logs
nodes

Inspiration

We all have that one friend who's always up to date on everything; you ask them about any topic and they just know. We wanted to build that friend. Someone you tell "catch me up on what's happening with X" and they go search articles, watch videos, and come back with a clean summary. Researchers, newsletter writers, content creators; they all spend hours doing this manually. We wanted to make it one click.

What it does

You give Newsloop a topic and a date range. It goes out, searches the web for articles and YouTube videos at the same time, reads the articles, watches the videos (via Reka Vision AI), checks if it found enough, and if not, goes back for more. Then it writes you a structured digest with sources. You can watch the whole process live in a web interface or run it from the terminal.

How we built it

LangGraph for the agent orchestration with parallel branches; text and video run at the same time. Tavily for web search, BeautifulSoup for article extraction, Reka Vision for video understanding, Pioneer AI for entity recognition, and OpenAI for the reasoning. The web interface is FastAPI with real-time streaming to a neo-brutalist UI. We used Claude Code for implementation.

Challenges we ran into

Getting two parallel branches to merge cleanly in LangGraph without one overwriting the other. Video indexing on Reka takes minutes while text finishes in seconds, so we had to design around that. YouTube discovery through Tavily often returns nothing for specific queries, so we added a broad fallback strategy. Real-time log streaming across threads was trickier than expected.

Accomplishments that we're proud of

The parallel architecture actually works; watching both branches run simultaneously and merge into one report is satisfying. The agent genuinely evaluates its own coverage and decides to loop back if it's not good enough. Both sponsor integrations (Reka and Pioneer) serve a real purpose in the pipeline, they're not bolted on.

What we learned

The hardest part of building agents is orchestration, not individual nodes. Handling failures gracefully, designing retry loops, merging parallel outputs that's where the real work is. We also learned that real-time visibility into what the agent is doing makes it dramatically more trustworthy than a black box.

What's next for Newsloop

More source types; podcasts, Twitter threads, Reddit, arxiv papers; each as a new parallel branch. Memory so it can track topics over time and flag what's new. Scheduled runs that send daily briefings by email. A model router that picks the best LLM per node instead of using one for everything.