The AI Agent Stack That Runs a Newsroom

There is a moment, around 4am, when News Beast publishes its first edition of the day, and not a single human being is awake to see it happen. No editor approved the headline. No journalist filed the story. By the time Malta wakes up, ten distinct editorial voices have already filed their pieces — Gabriel Fenech has dissected the latest political maneuver, Marcus Azzopardi has translated overnight market moves into something a nurse driving to her shift can actually use, and Dua Mifsud has already decided what last night's Love Island Malta episode secretly revealed. I built this. Not a team. One person, a stack of tools that don't talk to each other by default, and several months of forcing them to.

The mistake everyone makes first

The obvious approach to AI content is: one AI, one prompt, churn out articles. This is why almost all AI content reads the same — competent, hollow, interchangeable. News Beast runs ten separate author personas, each with an isolated source pool and a full character — biography, voice rules, private histories that shape every sentence without ever being directly referenced. None of this appears in the articles. All of it is in the system prompt. The reader feels the difference without knowing why.

The three-layer isolation problem

Ten AI writers pulling from overlapping news sources produces the same story written ten different ways on the same day. That's not variety — that's redundancy in a costume. News Beast solves this with three layers of isolation: Source isolation — Gabriel's RSS pool covers only Maltese local news. Marcus's covers only finance. They never see the same headline. Topic isolation — non-overlapping keyword matching per category. A story can't bleed into two writers' queues. URL blacklist — any used source URL is blacklisted for seven days, per category. The result: ten writers, twenty-two categories, forty editions a day, almost zero topic collision — not because the AI is clever enough to avoid repetition, but because the architecture makes repetition structurally impossible.

What actually happens at 4am

A Cloudflare Worker wakes on a cron schedule. It pulls fresh articles from each author's isolated RSS pool, scores them against that category's keywords, and hands the top candidates to Claude acting as chief editor — pick the eight best stories, enforce diversity, select a cover image. Then the writer model runs with the full character system prompt — sometimes two thousand words of biography, voice rules, banned phrases — plus one rule that matters more than anything else: you are not summarising sources, you are telling one story. After the draft, a second AI call almost nobody building "AI content" thinks to add: an editor's note. A different character reads the finished piece and reacts — agrees, pushes back, adds their own experience. Not a quality check. A newsroom behaving like a newsroom, automatically. The piece is stored — Supabase Postgres as source of truth, Cloudflare D1 as the edge replica — and IndexNow fires before the human reader even exists. Total time from trigger to published, indexed article: under thirty seconds.

The stack underneath

None of this exists in a vacuum — it sits on infrastructure most teams already have access to but rarely combine this way: The orchestration could just as easily run on N8N or Make.com instead of a raw Cloudflare cron — both handle the same scrape-score-generate-publish loop with a visual interface, which matters if you're not writing JavaScript by hand at 2am. Claude API does the actual writing and editorial review. Cloudflare serves the result from the edge in under 80ms, the difference between a newsroom and a newsroom nobody reads because it loads slowly.

The fallback nobody builds, and should

The most underrated decision in this system is the never-empty fallback — born from one bad morning when an API quota silently failed and a category showed "content not available" for six hours. Now if today's edition isn't ready, the system serves yesterday evening's. If that's missing, yesterday morning's. There is no version of News Beast where a category shows an empty state.

Why this matters beyond journalism

I didn't build this to prove a point about AI. Malta didn't have an AI-native newsroom, and someone had to be first. But the architecture underneath — isolated agent pools, layered deduplication, structured character prompts, automated peer review, never-empty fallback — works for any business trying to use AI agents for real output instead of demos. Most "AI content" is a single generic prompt against a single generic dataset. The fix isn't a smarter model — Claude, GPT, Gemini are all capable enough now. The fix is treating the system like an actual newsroom, then letting AI execute that structure at a speed no human team could match. It was never about the AI being good enough. It was about whether anyone bothered to build the newsroom around it. --- If you're trying to build something similar — an AI agent system that behaves like a team, not a chatbot wearing a content calendar — this is exactly what FreeMalta's Fractional CAIO service builds for companies who want it done properly the first time.