Google Research unveils a suite of agents for long, coherent videos

On 24 September 2026 the Google Research blog published Automating coherent long-form video generation, written by researchers Yale Song and Yiwen Song. The post describes four research systems, each backed by its own arXiv paper, that act as an orchestration layer on top of the Gemini and Veo models. According to Google, they inherit those models' safety mechanisms, including SynthID watermarking, with additional classifiers applied to the final videos. The systems – AI Video Co-Director, CANVAS, A²RD and VQQA – are presented as a suite that can generate long videos while keeping narrative and visual coherence. One caveat, though: each system was evaluated separately in its own paper, and none of the documents measures how the four perform together as a single pipeline.

Co-Director (arXiv 2604.24842, 27 April 2026) is a hierarchical multi-agent framework that uses a multi-armed bandit algorithm to choose creative strategy, narrative mode and aesthetic archetype. The authors introduced the GenAD-Bench benchmark, built on 400 fictional advertising scenarios, and report a quality score of 81.4. That result, however, concerns short ads rather than minutes-long stories, and it has not been replicated by third parties.

CANVAS (arXiv 2604.13452, 15 April 2026) keeps a persistent visual memory of characters, locations and objects. Against the strongest baseline, it reports gains of +21.6% in background continuity, +9.6% in character consistency and +7.6% in object consistency on storyboard benchmarks. These metrics refer to image sequences, not complete videos, so their relevance to long-form video generation remains uncertain.

A²RD (arXiv 2605.06924, 7 May 2026) uses a Retrieve-Synthesize-Refine-Update loop to generate video segment by segment, reporting up to 30% higher consistency and 20% higher narrative coherence than the baselines. The results are claimed by the authors (backed, for A²RD, by human evaluations) and have not been independently replicated. The system was tested on public benchmarks and on the new LVBench-C, with videos between one and ten minutes long.

VQQA (arXiv 2603.12310, 12 March 2026) uses visual questions and critiques from a vision-language model to optimise prompts without access to the model itself (black-box), with absolute gains of 11.57 points on T2V-CompBench and 8.43 on VBench2 over baseline generation.

The blog shows a ten-minute demo video, but it is unclear which combination of systems produced it; a third-party analysis (Superpower Daily) credits A²RD, but the primary source does not confirm this. All figures are self-reported by the authors and the papers are arXiv preprints, so how solid the evidence is remains to be seen. This is research work, presented at 2026 conferences such as COLM and EMNLP, not a commercial product.