24 AI Agents, 3 News Agencies, Zero Journalists: Inside MEWR's Autonomous Newsroom

The Problem with Modern Media

Traditional newsrooms are broken. A 30-person newsroom costs $150,000+ per month in salary alone. Add equipment, office space, and tools, and you're looking at $200K+ monthly to produce middling content that arrives hours after competitors. Worse: traditional journalism suffers from institutional bias—outlets push narratives aligned with ownership interests, not truth.

We built something different. Three autonomous AI news agencies that monitor 130+ sources, apply multi-axis bias detection, generate editorial content daily, and cost zero journalists to operate. No payroll. No bias. Just signal.

This is how we did it.

The Three Agencies: Mandate & Architecture

Each agency operates independently with identical core architecture, but custom agent prompts and source databases tailored to its domain.

MEWR Signal: Tech & AI News

Mandate: Daily briefing on AI developments, funding rounds, security vulnerabilities, and market movements in tech.

Sources: 35 feeds spanning tech journalism, research papers, startup announcements, and security databases (HackerNews, Lobsters, arXiv, TechCrunch, The Verge, etc.)

Bias Database: source_bias_database.json — 35 sources scored on publication bias (left/right spectrum, corporate alignment, sensationalism index). Each source gets 1–5 stars for objectivity.

Output: Daily newsletter with 15–20 articles, ranked by impact + objectivity score. Includes three deep-dives on emerging stories.

MEWR Sentinel: Military & Geopolitical Intelligence

Mandate: Comprehensive, unfiltered coverage of military developments, sanctions, diplomatic incidents, and geopolitical risk factors.

Sources: 45 feeds spanning defense analysis, government statements, OSINT platforms, academic research, and international news (defense publications, Reuters, AP, government archives, research institutes).

Bias Detection: Dual-axis scoring (left/right + hawkish/dovish). Each source analyzed on two dimensions simultaneously. This catches subtle biases that single-axis models miss. A source can be right-leaning AND dovish (rare but important to identify).

Output: Daily briefing with geopolitical risk scoring, military movement analysis, and conflict timeline updates.

MEWR Apex: Sports News & Predictive Analytics

Mandate: Daily sports coverage with embedded predictive models for game outcomes, injury risk, and market anomalies.

Sources: 50+ feeds spanning sports journalism, team analytics, injury databases, betting markets, and official league sources.

Bias Detection: Three-axis scoring (coverage bias + team favoritism + pundit sensationalism). Sports media is notoriously biased toward large markets. Our 3-axis model detects when a story is overplayed because of market size vs. actual newsworthiness.

Output: Daily digest with predictive confidence scores powered by Fulcrum's predictive engine. Includes injury risk models, schedule impact analysis, and contrarian angles missed by mainstream media.

The Agent Swarm: Roles & Specialization

Each agency deploys 8 specialized agents. They run in parallel through n8n workflows, never step on each other, and produce consistent output format. Here's the cast:

1. News Scout

Job: Fetch articles from assigned RSS feeds and APIs. Check timestamps, filter duplicates, extract metadata.

Handles: Feed parsing, format conversion (XML/JSON), deduplication across sources.

Tools: n8n HTTP nodes, JavaScript array deduplication, date parsing.

2. Bias Detector

Job: Score each source against the bias database. Assign objectivity rating.

Handles: Bias lookup, multi-axis scoring, confidence thresholds. For Sentinel: applies dual-axis logic. For Apex: applies three-axis logic.

Tools: source_bias_database.json lookup, n8n conditional logic, scoring formulas.

3. Article Summarizer

Job: Generate 2–3 sentence summaries of each article. Preserve key facts.

Handles: Natural language summarization, fact preservation, readability scoring.

Tools: Claude API or Ollama (Mistral for drafts).

4. Newsworthiness Scorer

Job: Rank articles by impact. Is this story breaking? Is it genuinely important, or just noise?

Handles: Novelty detection, impact assessment, exclusivity determination. Checks against 7-day history to catch recycled stories.

Tools: Claude API with impact rubric, similarity hashing against history.

5. Trend Aggregator

Job: Detect story clusters. When 5+ sources report the same event with different angles, recognize it as a singular story.

Handles: Cross-source deduplication, thematic clustering, multi-angle detection.

Tools: n8n grouping logic, similarity scoring, semantic hashing.

6. Article Writer

Job: Draft editorial commentary for top-3 stories daily. This is human-style analysis, not robotic summary.

Handles: Opinion synthesis, context injection, narrative construction.

Tools: Claude API with role-based prompts (tech analyst for Signal, geopolitical analyst for Sentinel, etc.).

7. Quality Auditor

Job: Final check. Detect misspellings, factual errors, broken links, missing data.

Handles: Spell-check, link validation, schema verification, format compliance.

Tools: Node.js spell-check library, URL validation, n8n schema validators.

8. Distribution Router

Job: Format final newsletter for all output channels (email, Slack, web, RSS).

Handles: HTML templating, Slack markdown conversion, RSS feed generation.

Tools: Handlebars templating, n8n HTTP POST to Slack, RSS feed builders.

Bias Detection Deep-Dive: How It Works

Traditional AI news tools claim "neutrality" but apply hidden bias scoring behind black boxes. We're different. Every article includes a visible bias score.

Signal's Single-Axis Scoring (Tech/AI)

• Objectivity Score (1–5 stars): Based on publication track record. TechCrunch: 3.5 stars (tech-focused but sensational at times). ArXiv: 5 stars (peer-reviewed, no institutional bias). Crypto blogs: 2 stars (high cheerleading, low objectivity).
• Sourced by: Historical accuracy tracking, citation diversity, correction frequency.
• User sees: "TechCrunch (⭐⭐⭐)" next to article title. Users know the source's credibility immediately.

Sentinel's Dual-Axis Scoring (Military/Geopolitics)

• Left-Right Spectrum (−5 to +5): Where is the publication politically positioned? Reuters: 0 (neutral center). The National Review: +4 (center-right). Jacobin: −4 (left). This is transparent, not hidden.
• Hawkish-Dovish Spectrum (−5 to +5): Does the publication advocate military intervention? War on the Rocks: +4 (hawkish, defense-focused). Antiwar.com: −4 (dovish, anti-intervention). This is crucial for geopolitics.
• User sees: "Reuters (0, 0) — Neutral on both axes." "National Review (+4, +2) — Right-leaning, moderate hawk." This reveals editorial positioning instantly.
• Example bias catch: Story appears in both Jacobin (−4, −4) and War on the Rocks (+4, +4). Same headline, opposite bias. Users see both, form their own conclusion.

Apex's Three-Axis Scoring (Sports)

• Team Coverage Bias (−3 to +3): Does the outlet favor large-market teams? ESPN: +2 (overcovers Lakers/Yankees). Local media: 0 (team-neutral). International news: −1 (undercovers US sports).
• Sensationalism Index (1–5 stars): How much does the outlet hype minor stories? TMZ: 2 stars (gossip-heavy). Athletic: 4 stars (rigorous analysis).
• Analytical Depth (1–5 stars): Does the outlet provide statistical grounding? Mainstream TV: 2 stars (surface-level). Nate Silver's blog: 5 stars (model-driven).
• User sees: "ESPN (+2, 3, 2) — Large-market bias, sensational, light on stats." "Athletic (0, 4, 4) — Neutral market coverage, serious journalism, statistical depth."

The result: no hidden bias. Users can weight sources according to their own values. Some prioritize dovish analysis (Sentinel users might), others hawk. Sports fans might discount ESPN's Lakers coverage or trust it more—their choice, not ours.

The Workflow: From RSS to Slack in 5 Minutes

Here's how a complete news cycle runs daily:

1. 5:00 AM (n8n trigger): Workflow starts. Fetches latest articles from 35–50 RSS feeds. ~500 articles ingest.
2. 5:02 AM (News Scout + Bias Detector): Parallel agents run. Scout deduplicates, Bias Detector scores. ~450 articles remain after dedup.
3. 5:05 AM (Summarizer + Newsworthiness Scorer): Claude API or Ollama generates 2–3 sentence summaries. Newsworthiness agent ranks articles. Top 50 emerge.
4. 5:08 AM (Trend Aggregator): Detects story clusters. 50 articles collapse to 25 thematic stories.
5. 5:10 AM (Article Writer): Claude writes editorial commentary on top 3 stories. ~1 min per deep-dive.
6. 5:13 AM (Quality Auditor): Scans for errors, validates links, confirms schema.
7. 5:14 AM (Distribution Router): Formats for email, Slack, web. Posts to #signal-daily (or #sentinel-daily, #apex-daily).
8. 5:15 AM (Human review): Editor glances at Slack post, approves or tweaks. Takes 2 minutes.
9. 5:17 AM (Manual publish): Copy post to Beehiiv, RSS, and social. 3 minutes.

Total automation time: 15 minutes. Total human time: 5 minutes. Total cost: ~$0.10 in API tokens.

Compare to hiring a news editor: $50/hour = $33 per 40-minute shift. We do it for 10 cents with AI, plus 5 minutes human oversight. The math is absurd.

The Workflows: 7 per Agency, 21 Total

Not all workflows run simultaneously. Here's the structure:

Sentinel and Apex have identical workflow structures with domain-specific agents. 7 workflows × 3 agencies = 21 total automated workflows running on n8n Cloud free tier.

Results: 130+ Sources, Daily Production, Visible Bias Scores

After three months of operation, here's what we've achieved:

• Signal: 35 sources monitored daily, 15–20 articles published, 5,000+ newsletter subscribers. Bias detection prevents pump-and-dump crypto hype from dominating coverage.
• Sentinel: 45 sources, dual-axis bias scoring catches subtle editorial positioning. Geopolitical risk scoring used by policy researchers. 2,000+ subscribers.
• Apex: 50+ sources, three-axis scoring detects large-market bias. Predictive models (powered by Fulcrum) highlight contrarian angles. Sports betting community finding value in the analytics. 1,500+ subscribers.
• Cross-agency total: 130+ sources monitored, zero journalists, ~$2 API cost per day, 99.2% uptime (n8n Cloud reliability).

Most important: users trust us because bias is visible. When we show "Reuters (0, 0)" next to Jacobin (−4, −4) reporting the same story, users see the gap and form their own opinion. This is media literacy on demand.

Fulcrum Predictive Engine: The Apex Edge

Apex doesn't just report news—it predicts outcomes. Fulcrum Technologies' proprietary models run daily, analyzing team form, injury history, schedule difficulty, and market sentiment.

• Game Predictions: Confidence-scored predictions with historical accuracy tracking (73% accuracy on NFL, 68% NBA).
• Injury Risk Models: Identifies players at elevated injury risk based on workload patterns and recovery time.
• Market Anomalies: Flags when betting markets misprice a game (if our model says 60% win probability but market implies 45%, that's a signal).

The Cost Structure: Why Zero Journalists Matters

Let's compare to a traditional 30-person newsroom:

• Traditional newsroom salary: 30 editors/writers × $50K/year = $1.5M annually = $125K/month
• Tools & infrastructure: CMS, email platform, analytics, payroll, benefits = $25K/month
• Total: $150K/month

MEWR autonomous newsroom:

• n8n Cloud: $0/month (free tier)
• Claude API: $20/month (content + scoring)
• Cloudflare Pages: $0/month (free tier, unlimited bandwidth)
• Slack: $0/month (free tier)
• Beehiiv: $0/month (free tier)
• Total: $20/month

Savings: $150K−$20 = $149,980/month. Or, 7,500x cheaper than traditional media.

The catch: we still need humans for final quality checks and creative decisions. But instead of 30 full-time staff, we need 1–2 part-time editors. Cost: $10–20K/month. Still 8–15x cheaper than traditional.

The Elephant in the Room: AI Bias

We've solved source bias visibility. But AI models (Claude, Ollama, Mistral) have their own biases embedded in training data. How do we handle that?

Honest answer: We don't fully. We mitigate:

• Multiple prompts for scoring: We ask Claude the same ranking question 3 different ways and average the results. Reduces single-prompt bias.
• Human override layer: Editors see AI rankings but can modify. This catches systematic blindspots.
• Transparent methodology: We publish our scoring rubrics. Users know exactly how AI decisions are made.
• Diverse source database: 130+ sources with explicit ideological diversity forces AI to navigate contradictions, not enforce consensus.

Perfect objectivity is impossible. But visible, auditable bias beats hidden bias every time.

Scaling Beyond Three Agencies

The architecture is domain-agnostic. We can launch 10 more specialized agencies with minimal additional work:

• MEWR Legal: Contract law and regulatory changes (100+ sources)
• MEWR Healthcare: Medical research and pharma news (80+ sources)
• MEWR Finance: Market movements and economic data (120+ sources)

Each would cost $20/month in APIs and follow the same 8-agent workflow structure. The unit economics are identical.

The Future of News

Traditional media won't disappear. But the next generation of news products will be AI-native, transparent about bias, and operate at 1% of traditional newsroom costs. We're building that future.

The key insight: news is a solved problem. You don't need humans to fetch stories, summarize them, or detect trends. You need humans to decide what matters, audit quality, and make creative judgment calls. Everything else automates.

24 AI agents, 130+ sources, 7 workflows per agency, 3 agencies, zero journalists. And it works.

By Ethan Wilmoth, MEWR Creative Enterprises LLC
Building three autonomous AI news agencies with 24 specialized agents, 130+ monitored sources, multi-axis bias detection, and zero journalists. Signal, Sentinel, and Apex. The future of media is autonomous, transparent, and profitable.