Case Study
A fully automated multi-channel content pipeline. From topic research to published video. 60 AI agents, 6 pipeline stages, 34 pieces per week across 5 channels. Total cost: $50/month.
A fully automated content production system that takes a topic and turns it into a published video with zero human intervention. Script writing, voice generation, video editing, quality review, thumbnail creation, publishing, and cross-posting all happen autonomously through 60 specialized AI agents.
It runs 5 channels in parallel: finance, AI tools, sleep content, language learning, and narrative entertainment. Channel selection was driven by RPM analysis (revenue per thousand views). Each channel has its own agent configuration, content templates, and quality benchmarks.
Topic research agents analyze trending topics, search volume, competition density, and RPM potential. They generate a content brief with hook, outline, target keywords, and estimated performance.
Claude API generates the full script from the content brief. Channel-specific writing agents apply the right tone, pacing, and structure. A finance explainer sounds different from a sleep meditation.
Edge TTS converts the script into natural-sounding narration. Each channel has a consistent voice profile. Pacing, pauses, and emphasis are controlled through SSML markup in the script.
Auto-editing agents assemble the video, matching stock footage to script segments, adding transitions, text overlays, and background music. Each cut is timed to the narration.
Multi-agent review. One agent checks script quality, another reviews audio sync, a third evaluates thumbnail appeal, and a fourth runs the whole piece against channel benchmarks. Pieces that don't pass get sent back to the relevant stage.
Flux AI generates the thumbnail. Metadata agents write titles, descriptions, and tags optimized for each platform. Publishing agents upload and cross-post across platforms with proper formatting for each.
The brain. Handles topic research, script generation, quality review, and metadata creation. Each agent has a specific system prompt, retry logic, and quality threshold before passing to the next stage.
Text-to-speech engine for narration. Free tier, high quality, SSML support for controlling prosody. Different voice profiles per channel for brand consistency.
Thumbnail generation. Creates eye-catching, platform-optimized thumbnails from the content brief. Style-matched to each channel's visual identity.
60 agents total, each with a single responsibility. Research agents, writing agents, voice agents, editing agents, review agents, publishing agents. Each stage has its own retry logic and quality checks.
The quality gate is the key. Most automated content pipelines produce volume but not quality. This system has a multi-agent review stage where different agents evaluate different dimensions: script quality, audio sync, visual appeal, metadata optimization. Content that doesn't meet the threshold gets recycled back to the relevant stage, not published anyway.
The economics are the other differentiator. 34 pieces per week across 5 channels for $50/month. That's roughly $0.21 per published piece. A human content team producing the same volume would cost $15,000+/month. The ROI math isn't close.