Assistant or Agent: Building the Same Content Machine Three Ways

Anatoly Klavdienko, responsible for AI agent development at AlpinaGPT and Alpina Digital, automated production of 5-10 short videos per week for an AI/tech Telegram channel. He first used ChatGPT as a chat assistant, which took about 4 hours per week for 10 videos but involved manual copying and assembly. Then he built an n8n workflow, spending four days on setup and debugging, but found the visual editor painful for iterations, lacking versioning and proper debugging. Finally, he exported the n8n workflow as JSON, fed it to Claude Code, and got a working Python pipeline in one evening. The final agent consists of 2074 lines of production code in 19 files, with three pipelines: D (sourcing from 12 RSS feeds via GPT-4o), C (generation via HeyGen with 11 look IDs, Whisper and FFmpeg post-processing), and A (publishing to YouTube, TikTok, and Instagram Reels). Real timings from pipeline.log show sourcing takes about 80-84 seconds, generation 6-8 minutes. When the SubMagic API started failing on April 20, 2026, the author replaced it with a custom post-processing service in one day. He concludes that code agents are now more economical than visual builders, offering better control and debuggability.