An animated telenovela, produced by agents

A Spanish-language micro-telenovela — one-minute episodes, cliffhangers, a cast with its own canon — published as vertical video. It found a real panhispanic audience fast, and it exists to answer a serious question: how much of a content operation can agents actually run?

Seven specialized agents, each with its own instructions, tools and persistent memory. I don't operate the pipeline — I designed it, and I audit it.

• 01 · Writerwrites each episode from the season arc: hook, beats, cliffhanger.
• 02 · Promptsturns the script into per-clip technical prompts under a rulebook per model.
• 03 · Validatorinspects every generated frame against each character's canonical sheet — verdict by dimension, before money is spent on video.
• 04 · Editorassembles the episode locally: cuts, music bed, karaoke subtitles, watermark.
• 05 · Publisherproduces the publishing plan: captions, hashtags, platform cascade.
• 06 · Datapulls performance data and writes the postmortem with hypotheses.
• 07 · Strategistreads the data and proposes the next strategic play.

Orchestrated with n8n calling the Claude Agent SDK; image and video generated on Vertex AI (Veo); assembly runs fully local with ffmpeg and Whisper — that half of the pipeline costs zero. Context lives in a written canon the agents load selectively, so any agent can be replaced without losing the show's memory.

Nothing spends without passing a gate: frames are validated cheap before video is generated expensive, every generation call carries a budget cap, and a human GO is required to run. The validator was rebuilt after failing — the lesson (a validator fails on input format, not capability) is now part of how I build every LLM check.