Google Research has introduced an AI video co-director, a multi-agent framework designed to generate temporally consistent long-form video narratives. The approach is an orchestration layer on top of Gemini and Veo, and it inherits native safety features such as SynthID watermarking. The researchers argue that current chained pipelines suffer from semantic drift, cascading failures, feature drift, and content collapse, often requiring exhaustive manual intervention.
Instead of relying on rigid linear prompt chains, the framework formalizes storytelling as a global optimization problem. A multi-armed bandit algorithm selects creative configurations across three dimensions: creative strategy, narrative mode, and aesthetic archetype. Those choices are injected into sub-agents that handle pre-production, keyframes, video, and audio, while a multimodal LLM judge critiques the compiled cut and feeds a factored reward signal back to the bandit for iterative refinement.
The work is part of a suite of frameworks — Co-Director, CANVAS, A²RD, and VQQA — that automate repetitive orchestration tasks and closed-loop visual refinement. In evaluations, the authors report substantial gains in multi-shot narrative consistency and character persistence, generating minutes-long videos while mitigating visual drift and error propagation. The source is a single Google Research blog post; no independent results or comparisons are included.