Shoestring Studios

AI-generated explainer content typically restates its source: it rarely draws on outside theory or real-world cases, weights every point equally, and goes unreviewed. As a result, it falls short of what a human expert provides, which is not a summary of the page but an account of what it means. This project, Shoestring Studios, addresses that gap by giving AI what an expert relies on: a team. Shoestring Studios is a video studio staffed by twelve AI employees, each a ReAct agent running on Muse Spark Open. Given a PDF and a brief taken through a conversational order, a seven-member storyboard team works in sequence to set the concept, research theory and dated real cases, plan scenes, write and voice the narration, and draw a storyboard. A five-member visual team then operates on every part in parallel, sourcing the document's own figures, generating images with Muse Image, animating data, verifying visuals with vision, and rendering the final MP4. A multi-agent pipeline of this size is constrained less by model capability than by data flow. In an initial design in which every agent received all upstream material, a single agent's first prompt reached nearly 250,000 tokens. To resolve this, a new framework, Nextop Agentic Workers (NAW), was developed during the hackathon. NAW models each agent as a block that declares only its inputs and outputs, without a coordinating manager agent; splits each output and routes the pieces only to the agents that consume them, with the entire topology defined in a single editable flow file; and provides shared LightRAG-indexed stores from which each agent retrieves only the passage, figure, or table it requires. Under NAW, the same agent's first prompt fell to approximately 8,000 tokens. Applied to a book chapter, the system produced a 10-minute-48-second video comprising 43 cuts, 31 photographs, 6 generated images, and one table, with picture and narration of equal length and no silent gaps.