Curate AI news and draft social posts with a local agent newsroom

Kiran Beegala runs a chain of small agents on a local 8B model that ranked about 2,900 AI news items in ten days and drafts Bluesky and LinkedIn posts that wait for one Telegram tap.

What they did

Kiran Beegala got tired of skimming eight tabs of AI news each morning and built a newsroom that runs on a 16 GB MacBook with no cloud model keys. All reasoning goes through qwen3:8b on Ollama. A chain of small agents collects items from about 21 public sources (arXiv, GitHub, Hugging Face, Hacker News, tech blogs), removes duplicates, scores them, researches the top few, then summarises, drafts and critiques. It writes a daily Bluesky digest with image cards, a weekly LinkedIn carousel, LinkedIn posts for big breaking news and a short narrated reel each day. The agents talk over A2A, use MCP for tools, keep state in SQLite and run on a schedule. Every draft goes to Telegram, where the author approves, asks for changes or skips.

How it went

In its first ten days the pipeline took in about 2,900 items. LinkedIn and Bluesky posts have gone live after approval, including a five-page LinkedIn deck built from a GitHub repo. The reels are built daily and reviewed, but posting them to Instagram is still manual. The author gives no time-saved or engagement numbers, and the post has no comments yet.

Worth knowing

The author does not trust the 8B model with facts. Code checks every figure in a draft against the source material, flags filler phrases for a rewrite, and inserts all links itself so the model never writes a URL. A small web UI replays each run so a bad post can be traced to its exact prompt. The code is not shared.

Try it

With Custom agent· 7 parts to fill in

Help me build a small news pipeline that runs on [my laptop] with [my local model, e.g. qwen3:8b on Ollama]. Each morning, pull items from [list of sources], remove duplicates, score them for [my audience], and research the top [number]. Draft one post per platform for [Bluesky, LinkedIn or my platforms]. Before any draft is shown to me, check in code that every number appears in the source text, flag filler phrases for a rewrite, and add the real links yourself rather than letting the model write them. Send each draft to me on [Telegram or my chat app] with Approve, Change and Skip options. Publish nothing until I approve that exact post.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related