Building SnowThere: An Autonomous Content Pipeline with a Three-Agent Editorial Panel

SnowThere is a fully autonomous content pipeline that publishes family ski resort guides across 116 resorts in 16 countries with zero human editors. A three-agent editorial panel reviews every piece before publication, running on a $5/day budget. Here is the architecture.

The Problem: Content at Scale Is an Editorial Problem

Building a comprehensive family ski resort directory sounds like a content problem. Research each resort, write a guide, publish it. But at 116 resorts across 16 countries, it's really an editorial problem. How do you maintain consistent quality, accurate facts, and a coherent voice across hundreds of pages of content without an editorial team?

The traditional answer: hire writers, editors, and fact-checkers. Budget $50-100 per article. Wait weeks for coverage. Watch it go stale within months as lift prices change, new childcare facilities open, and trail maps get updated.

We wanted to see if a different architecture was possible. Not AI-assisted content with human editors. Fully autonomous content with AI editors.

The Architecture: Research, Generate, Review, Publish

SnowThere's content pipeline runs on a daily cron job hosted on Railway. Every day, it follows the same four-phase cycle:

Phase 1: Research

Three research agents gather data about each resort from different sources:

• Semantic search APIs for finding recent articles, reviews, and updates
• Brave Search for structured data (lift counts, elevation, trail stats)
• Tavily for real-time information (weather conditions, operational status)

The agents run in parallel. Each returns structured data that gets merged into a comprehensive research dossier. Conflicting facts get flagged for the editorial panel.

Phase 2: Generation

Claude takes the research dossier and generates a full resort guide. The prompt includes our voice profile, structural requirements, and content standards. Every guide covers the same sections: overview, terrain, family amenities, childcare, beginner friendliness, getting there, and practical tips.

Consistency matters here. A family reading the Zermatt guide and then the Whistler guide should feel like the same trusted source wrote both.

Phase 3: The Three-Agent Editorial Panel

This is the part that makes SnowThere different from "AI-generated content." Every piece of content goes through a voting panel of three specialized agents before publication:

• TrustGuard checks factual accuracy. Are the lift counts correct? Do the prices match current data? Are safety warnings appropriate? TrustGuard cross-references claims against the research sources and flags anything it can't verify.

• FamilyValue evaluates family-friendliness. Is the childcare section detailed enough? Are beginner terrain options clearly described? Would a parent reading this have enough information to make a booking decision? FamilyValue has strict standards for what constitutes a useful family guide vs. a generic resort overview.

• VoiceCoach enforces editorial voice. Is the tone warm and helpful? Is the reading level accessible? Does the content avoid the breathless hype that plagues travel writing? VoiceCoach ensures every guide sounds like it was written by a knowledgeable friend, not a marketing department.

Each agent votes: APPROVE, REVISE, or REJECT. A 2/3 majority is required for publication. If any agent votes REVISE, the content goes back to generation with specific feedback. If any agent votes REJECT, the content is held for manual review (this has happened twice in hundreds of articles).

Phase 4: Publication

Approved content gets published through Next.js ISR (Incremental Static Regeneration). The site rebuilds only the pages that changed. Fresh content goes live within minutes of approval.

The Budget: $5 Per Day

The entire pipeline runs on approximately $5 per day. That covers:

• Claude API calls for generation and all three editorial agents
• Search API calls (semantic search, Brave, Tavily) for research
• Railway hosting for the cron job
• Supabase for data storage

Compare that to the traditional approach: 3-5 writers at $50-100 per article, an editor at $30-50 per article, and a fact-checker for high-stakes claims. For 116 resorts with quarterly updates, you're looking at $30,000-$50,000 per year in content costs. The autonomous pipeline costs roughly $1,800 per year.

Why Human Review Still Matters

Here's the counterintuitive learning: building a fully autonomous system made us more respectful of human judgment, not less.

The three-agent panel catches most quality issues. But "most" isn't "all." Twice, TrustGuard approved content where a resort had changed ownership and the new operator hadn't updated their website yet. The research sources all agreed on outdated information. A human would have caught the discrepancy from context clues the agents missed.

We handle this with a simple rule: any content that mentions safety-critical information (avalanche conditions, childcare licensing, medical facilities) gets a human spot-check before publication. It adds maybe 30 minutes per week to an otherwise autonomous process.

Fully autonomous doesn't mean fully unsupervised. It means humans review exceptions instead of reviewing everything.

The Results

The numbers after months of operation:

• 116 resorts published across 16 countries
• 2,000+ families matched to resorts through the Snow Quiz
• $5/day total pipeline operating cost
• 0 editors needed for ongoing content production
• 3-agent panel with 2/3 majority vote on every publication

The content quality holds up. Families consistently report that the guides are detailed, accurate, and useful for booking decisions. The Snow Quiz (a matching algorithm, not part of the content pipeline) has matched over 2,000 families to their ideal resort.

SnowThere is live at snowthere.com.

What This Pattern Means for Other Industries

The three-agent editorial panel is a pattern, not a one-off. Any business that produces content at scale can use this architecture:

• Real estate: Property descriptions with a FactChecker, BuyerValue, and BrandVoice panel
• E-commerce: Product descriptions with an AccuracyGuard, CustomerFocus, and ToneCoach panel
• Healthcare: Patient education content with a MedicalAccuracy, Accessibility, and Empathy panel

The specific agents change. The architecture doesn't: research, generate, multi-agent review, publish.

Build Your Own Autonomous Pipeline

Whether it's content, reports, or documentation, if your team produces high volumes of structured content, this pattern probably applies. Describe your content bottleneck and get a free Automation Blueprint that maps what an autonomous pipeline would look like for your use case. Takes 60 seconds.