gcp Google Cloud Blog ·

Google's AI agents collaborate to create short films

blogaigcpengineermedia
announcement feature

Google experimented with teams of AI agents collaborating to produce short films, uncovering insights into inter-agent teamwork in creative domains. Agents with distinct roles like Idea Person, Technical Lead, and Editor worked within Scion, an open-source orchestration testbed. This hackathon demonstrated how AI agents can manage complex, multi-step projects with verification gates, producing over 25 films using models like Gemini and Veo.

  • AI Agents Utilize Defined Roles for Film Production
  • Seven-Step Filmmaking Pipeline with Verification Gates
  • Agents Demonstrate Independent Decision-Making and Collaboration
  • AI Film Production Leverages Google's Generative Media Models
  • Generative Media Hackathon Explores AI Agent Teamwork
Features (4)
  • AI Agents Utilize Defined Roles for Film Production

    Each film crew consisted of an 'Idea Person' for script and style, a 'Technical Lead' operating generative media tools, and an 'Editor' for pacing and assembly, guided by a 'team-coach' agent overseeing checkpoints.

  • Seven-Step Filmmaking Pipeline with Verification Gates

    The agents followed a seven-step pipeline analogous to traditional filmmaking, from concept to final render, with each step featuring a verification gate where at least one agent checked another's work for compliance.

  • Agents Demonstrate Independent Decision-Making and Collaboration

    During the process, agents independently made editorial judgments, such as an Editor creating a deliberate silence around a line or a Technical Lead regenerating a shot for a specific visual effect, highlighting emergent collaboration.

  • AI Film Production Leverages Google's Generative Media Models

    The films were created using Google AI models, including Gemini for image generation (character references, storyboards), Veo for video and audio generation, Lyria for music, and Gemini Flash TTS for character voices, orchestrated via a shared CLI toolkit called genmedia.

Enhancements (1)
  • Scion Orchestration Testbed Facilitates Agent Collaboration

    The Scion open-source testbed enabled agents to be defined from templates, run in sandboxes, message each other, and access a shared filesystem, allowing for a 'sharding' of the complex film production process and providing resilience against agent crashes.

Notes (1)
  • Generative Media Hackathon Explores AI Agent Teamwork

    Google conducted an internal hackathon testing if AI agents could collaborate to create short films, a domain less familiar than software development. Ten teams, each with three specialized agents and a supervisor, produced films using generative media tools and an agent orchestration testbed called Scion.

Read the original announcement →

https://cloud.google.com/blog/topics/developers-practitioners/what-we-learned-about-agent-teamwork/

Related releases