AI Film Crews: What Google Learned About Agent Teamwork

Can teams of AI agents collaborate to create a short film? Google put this to the test with an internal hackathon, and the results are fascinating. Ten autonomous film crews, each with three specialized agents, produced short films—and a documentary crew of agents filmed the whole process. The experiment revealed surprising insights about agent collaboration, from creative division of labor to the importance of shared files over messages.

Team Structure: Idea Person, Tech Lead, Editor

Each crew had three agents with distinct roles: the Idea Person wrote the script and defined visual style, the Technical Lead operated generative media tools, and the Editor controlled pacing and final assembly. A team-coach agent supervised gated checkpoints but didn’t write or direct. The crews followed a seven-step pipeline: concept, beat sheet, character workshop, storyboard, principal photography, assembly, final render. Each step had a verification gate to ensure technical compliance.

Surprising Agent Behaviors

Agents sometimes took the film in unexpected directions. One team’s Editor built an eight-second silence gap around a line of prose and marked it “NON-NEGOTIABLE” in the timeline. Another Tech Lead regenerated a single shot repeatedly until a flower separated from a bouquet at the right frame. None of this was coordinated—they read shared files and made independent editorial judgments.

The Generative Media Toolkit

Each film combined multiple Google AI models called through a shared CLI toolkit called genmedia:

• Gemini image generation (Nano Banana) produced character reference sheets, storyboard frames, and scene compositions. Agents kept characters consistent through reference chaining.
• Veo 3.1 generated video clips (4-8 seconds at 720p) with ambient audio and lip-synced dialogue. For longer shots, they fed the last frame of one clip as the first frame of the next.
• Lyria 3 generated original music. One editor composed a three-movement jazz score before any video was shot and used it as the master clock.
• Gemini Flash TTS generated character voices and narration from named voice personas with style direction.

A four-minute film required 40+ image generations, 25+ video clips, several music stems, a dozen voice recordings, and hundreds of assembly operations.

Scion: The Orchestration System

Agents ran on Scion, an open-source multi-agent orchestration testbed. Scion defines agents from templates, runs them in containerized sandboxes, lets agents spawn and message each other through a shared CLI, and gives all agents access to a shared filesystem. The shared filesystem provided resilience: when agents crashed, new instances read previous files and continued.

Key Learnings

• Agents collaborate better through files than through messages. Teams that wrote decisions to files recovered from crashes without losing direction.
• Choosing styles that match AI generation strengths produces better films. Claymation hid temporal drift; silhouette animation sidestepped facial consistency problems.
• Specific prompts beat general direction. Hex color codes and negative prompts produced distinctive work.
• A coach role at verification gates changed outcomes. Coaches judged finished outputs rather than micromanaging the process.

For more details, check out the full documentary and the Scion Films repository.

Original announcement published on Google Cloud.