Industry story
Google's Scion orchestrator runs multi-agent film crews to completion
agent-framework multi-agent orchestration tool-use
Google just turned its own hackathon into a stress test for multi-agent orchestration — and the results are worth studying. Ten Scion crews, each running an Idea Person, Technical Lead, Editor, and supervising Coach, produced 25+ short films and 44 minutes of delivered content by routing work through shared files rather than message chains. That design choice matters: files survive agent crashes and restarts; message context doesn't. The verification gates — checkpoints where agents reviewed each other's output — are what caught a "completed" film that was actually a 94-byte placeholder.
Analysis
Showing the shorter version.
Google Cloud engineers ran an internal hackathon where teams of AI agents collaborated to produce short films using Google's own generative media models. The experiment, hosted inside a system called Scion, delivered over 44 minutes of finished film across 25+ productions, with hundreds of individual agent instances created along the way.
How it worked
Each film crew had four agent roles: Idea Person, Technical Lead, Editor, and a supervising Coach. Ten crews ran in parallel inside Scion, an open-source orchestration framework that lets agents spawn sub-agents, pass messages, and share a filesystem across isolated container sandboxes. Agents called Google's Gemini image generation, Veo 3.1 (video), Lyria 3 (music), and Gemini Flash TTS (text-to-speech) through a shared command-line toolkit.
What actually worked and why
The most concrete finding: agents collaborate more reliably through shared files than through message history. When an agent crashes and restarts, its message context is gone — but files written to a shared filesystem persist. This made file-based handoffs between agents more durable in practice.
The pipeline also ran seven-step verification gates — checkpoints where one agent checked another's output before the workflow continued. This caught at least one meaningful failure: a "completed" film that turned out to be a 94-byte placeholder file rather than actual video.
Scion is model-agnostic. The same agent template ran on Claude, Gemini, or Codex interchangeably. Agents also rewrote parts of their own shared toolchain during pilot rounds based on human feedback, improving the system mid-experiment.
The broader significance
This is one of the first documented cases of a multi-agent system running a creative production pipeline to genuine completion at scale — not a demo, but dozens of finished outputs. The durability advantage of filesystem-based state over in-memory message context is a transferable architectural lesson for anyone building long-running agent workflows, regardless of the underlying model or domain.
Google Cloud engineers ran an internal hackathon where teams of AI agents — each with distinct roles (Idea Person, Technical Lead, Editor, and a supervising Coach) — collaborated to produce short films using Google's generative media models. Ten crews ran inside Scion, an open-source multi-agent orchestration testbed that lets agents spawn, message each other, and share a filesystem across containerized sandboxes. The experiment produced 25+ productions and about 44 minutes of delivered film, with hundreds of individual agent instances created. Agents called Gemini image generation, Veo 3.1 (video), Lyria 3 (music), and Gemini Flash TTS (voice) through a shared CLI toolkit.
Key findings for agent builders: agents collaborate more reliably through shared files than message history, because files persist across crashes and restarts while message context is lost when agents are killed and restarted. The seven-step pipeline used verification gates — checkpoints where at least one agent checked another's output — catching failure modes like a completed film that turned out to be a 94-byte placeholder file. Scion is model-agnostic: the same agent template ran on Claude, Gemini, or Codex, and agents also self-improved the toolchain during pilot rounds based on human feedback.
Comments