# What 10 autonomous film crews taught us about agent teamwork > Source: > Published: 2026-07-16 16:00:00+00:00 Can teams of AI agents collaborate to create a short film? As part of an internal Google generative media hackathon, we put this question to the test – specifically, to uncover whether AI agents could work collaboratively in a domain less innately familiar than software development. We gave each crew three agents with distinct roles and had them collaborate through messages and shared files under their own agent-only hackathon. Agents ran inside [Scion](http://goo.gle/scion), an open source agent orchestration testbed. Unlike code or text, media and composition are less familiar subject matter for AI agents, so this experiment taught us about how agents can collaborate with checks and gates to see projects through to an end. Ten crews each produced a short film. A separate agent-staffed documentary crew "filmed" the process. That documentary itself became the medaling hackathon submission. The result? Hundreds of individual agent instances were created over the project. 25+ total productions across pilot rounds and competition. About 44 minutes of delivered film. Human feedback on the output fed back into a continuous improvement loops with the agent generated tooling. Here are two examples of agent generated short films: Each crew had three agents. The **Idea Person** wrote the script and defined the visual style. The **Technical Lead** operated the generative media tools. The **Editor** controlled pacing and final assembly. A team-coach agent supervised gated checkpoints but didn't write or direct. The **Idea Person** generated three starter ideas. Then, the team assessed the ideas from their role's POV: would this be generated well with generative media? Would it be complex to edit? Then, they pitched the idea among other teams in the hackathon, so that a team could adjust or pivot. For example, if three teams all picked a sci-fi space battle, then it would not make a good competitive entry. A **Coordinator agent** scheduled the competition, running two teams at a time across five waves. The event ran about 21 hours. The crews followed a seven-step pipeline modeled on the fundamentals of traditional filmmaking: concept, beat sheet, character workshop, storyboard, principal photography, assembly, final render. Each step had a verification gate, ensuring that at least one agent checked another agent's work for technical compliance (such as resolution, or timing). In an early pilot, one team reported a completed film that turned out to be a 94-byte placeholder file. As it turns out, agents can be convincing about having finished work they haven't done. While surprising (and sometimes even amusing), we uncovered other ways the agents took the film in their own direction. For example, the agents divided labor on their own in ways we didn't expect. On one team, the Idea Person wrote a line of prose in the first draft. The Editor, independently, built an eight-second silence gap around that line and marked it "NON-NEGOTIABLE" in the timeline. The Tech Lead regenerated a single shot repeatedly until a flower separated from a bouquet at the right frame. None of them coordinated this. They read the shared files and made independent editorial judgments. This process around teamwork and tool use was co-developed with agents during the pilot-phase. During this phase, agent teams created videos which received human feedback, such as audio collisions and levels, inconsistent characters, hard to follow story or narration. This feedback, combined with agent-authored retrospectives for each pilot was used to restructure not only the playbook and guides that instructed future teams through the process, but the agents also built and revised a custom media toolchain that combined golang CLIs with python batch automation. 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. The agents ran on [ Scion](http://goo.gle/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. **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." You can see the full documentary [here](https://youtu.be/WpnChAr_FDc), and learn more about the [Scion Framework](http://goo.gle/scion), and how it was [used in the hackathon](https://github.com/ptone/scion-films).