The generative media models
Each film combined multiple Google AI models. The agents called them through a shared CLI toolkit called genmedia:
Gemini image generation (Nano Banana) produced character reference sheets, storyboard frames, and scene compositions. The agents kept characters visually consistent across a film through reference chaining: they generated headshots first, then used those as input for body sheets, then used body sheets as input for scene tests. Each generation call included these accumulated references as anchors.
Veo 3.1 generated the video. Clips run four to eight seconds at 720p. The agents chose different generation modes depending on the shot: text-to-video for simple compositions, image-to-video for shots anchored to storyboard frames, frame interpolation when they needed a precise start and end frame. For shots longer than eight seconds, they fed the last frame of one clip as the first frame of the next.
Veo 3.1 also generates audio inside each clip: ambient sound, room tone, and lip-synced character dialogue. One team (Lambda) built their film around this capability. They structured the script like a musical score with movement markings (Allegretto, Accelerando, Adagio) because the sync between generated speech and lip movement gave pauses real weight.
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 for the production. Teams also coerced Lyria into producing sound effects by framing prompts as “soundscapes.”
Gemini Flash TTS generated character voices and narration from named voice personas with style direction (“world-weary narrator, slow measured pace”). TTS pacing was hard to predict. One team’s narrator delivered at 108 words per minute instead of the planned 130, blowing out the runtime by a full minute. A different team had a similar problem but decided the slow pace fit their character, a 68-year-old projectionist.
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
The agents ran on Scion, an open-source multi-agent orchestration testbed. Scion defines agents from templates (persona, instructions, skills, tools), runs them in containerized sandboxes, lets agents spawn and message other agents through a shared CLI, wakes agents through event-driven notifications, and gives all agents in a project access to a shared filesystem.
Messages and notifications allowed collaboration around a shared workflow. At different points in the process, different agents brought their focused contribution to that stage. Fundamentally this allowed for “sharding” the complex process across multiple context windows. Some of these were long lived, some short lived. Combinations of different models and harnesses were used as Scion is model and harness agnostic. The same agent template runs on Claude, Gemini, or Codex.
The shared filesystem provided resilience. Agents crash, run out of context window, and get restarted by the system. The files they write persist. When one team’s editor crashed during final assembly, the Tech Lead opened the editor’s timeline plan, read it, and finished the job. The coordinator restarted the documentary producer agent multiple times across the project. Each new instance read the previous one’s files and continued.
Some of what we learned
Agents collaborate better through files than through messages. Teams that wrote down their decisions (which visual keywords go in prompts, where shots sit on the timeline, what instruments to ban from the score) recovered from crashes without losing direction. Teams that kept decisions in message history lost them when agents restarted. The effective combination was to pass messages containing file-paths.
Choosing styles that match AI generation strengths produces better films. Teams chose claymation because its wobble made temporal drift invisible. They chose silhouette animation because it sidestepped facial consistency problems. One team couldn’t generate a kiss because a safety filter blocked it. They showed two shadows merging on a wall instead. Their coach called it the strongest shot in the film.
Specific prompts beat general direction. The default output from video generation is moody cinematic noir. The teams that made distinctive work specified hex color codes rather than color names, listed banned instruments, and wrote negative prompts ruling out unwanted aesthetics. “Make it warm” produced generic results. “#F4A261, no string instruments, no lens flare” did not.
A coach role at verification gates changed outcomes. The coach could observe the full production but could only intervene at step boundaries. That constraint forced coaches to judge finished outputs rather than micromanage the process. One coach described the dynamic: “It’s a room full of specialists who can each do one thing at superhuman speed, but none of them can taste the soup.”
Learn more
You can see the full documentary here, and learn more about the Scion Framework, and how it was used in the hackathon.
