Google researchers have built a multi-agent framework that generates minutes-long video while explicitly tracking characters, locations and objects from shot to shot. The system sits above Gemini and Veo, coordinating storyboards, keyframes, motion and audio instead of chaining independently written prompts.
Long AI videos often suffer from semantic drift, where clothes, faces or scenery change between clips, and from cascading failures caused by an early bad asset. Google's approach treats the full narrative as an optimization problem. An orchestrator chooses a creative strategy, story structure and visual style; production agents turn that plan into media; and a multimodal model critiques the assembled result so another generation loop can refine it.
A related component called CANVAS maintains a persistent visual memory of entities and environments. It retrieves earlier visual anchors when a character or place returns, giving later shots a concrete reference rather than relying on text alone. The broader suite also includes research on automated refinement and video question answering.
The team reports gains in character persistence and multi-shot consistency, but the announcement does not turn the framework into a generally available video editor. It is research scheduled for venues including COLM and EMNLP 2026. Generated images, video and audio inherit Veo's SynthID watermarking, while Google says additional classifiers may still be needed for interactions that only become unsafe in the completed sequence.