Google’s multi-agent stack targets coherent long-form video

Google Research has introduced a four-part framework that coordinates Gemini and Veo to generate coherent, minutes-long video. The work targets continuity failures across shots, including altered clothing, misplaced props, changing room geometry, and early asset errors that spread through later scenes. Google outlines the architecture in a research announcement and four accompanying papers.

Current video generators can produce photorealistic clips lasting several seconds. Assembling those clips into a narrative requires consistent characters, environments, objects, and story state across many generation calls. Google’s framework adds planning, persistent visual memory, evaluation, and revision around the underlying models.

Four systems divide the work

Google describes four related research systems, each addressing a separate part of the production pipeline. No downloadable product currently packages the full stack.

Small prompt errors spread

Long-form generation magnifies mistakes because every stage depends on assets and instructions produced upstream. A malformed keyframe can distort later motion, while an incomplete character description can change clothing or facial details across cuts. Independent prompts also lack a shared record of what has already happened in the story.

The team frames diagnosis as a credit-assignment problem. When the final cut fails, the system must identify which earlier decision caused the defect. Google groups the visible failures into two broad categories: feature drift, where characters or environments change unintentionally, and content collapse, where the narrative stops making meaningful progress.

The orchestrator searches creative options

The Co-Director paper places an Orchestrator Agent above the production pipeline. It uses a multi-armed bandit, an algorithm that allocates trials among competing options according to previous rewards, to explore three dimensions:

• Creative strategy: the intended message or objective
• Narrative mode: the structure used to develop the story
• Aesthetic archetype: the visual style and tone

The selected combination becomes shared guidance for every downstream agent. Production then moves through a defined sequence:

1. A Pre-Production Agent creates the storyboard.
2. A Keyframe Agent establishes characters, objects, and locations.
3. A Video Agent generates motion for each shot.
4. An Audio Agent adds narration, dialogue, and music.
5. A multimodal model judges the assembled cut and returns separate reward signals to the orchestrator.

The orchestrator uses those scores to allocate later trials toward more successful choices while preserving some exploration. Media generated through Veo retains Google’s SynthID watermarking.

CANVAS stores the story’s visual state

CANVAS, short for Continuity-Aware Narratives via Visual Agentic Storyboarding, maintains structured records for characters, locations, and object states as a narrative develops. When a setting or character returns after a cutaway, the system retrieves earlier visual anchors and incorporates them into the next generation step.

Google’s museum-heist example tests recurring details such as a thief’s cap, a gemstone, and the geometry of exhibition rooms. In the published comparison, a standalone Gemini-3.1-Pro pipeline changed props and room layouts, while the AutoStudio baseline dropped character details across cuts. CANVAS preserved more of the established identity and spatial structure.

A²RD extends synthesis to 10 minutes

A²RD generates video one segment at a time and is designed for sequences lasting up to 10 minutes. Every segment passes through a loop that retrieves relevant material from multimodal memory, synthesizes the next clip, refines it, and updates the stored state.

Its controller chooses between two generation modes. Extrapolation introduces new action and advances the plot. Interpolation reconnects the current segment to established characters, objects, and environments. Switching between these modes lets the system introduce change while preserving visual anchors across long gaps.

VQQA turns critiques into prompt updates

VQQA, or Video Quality Question Answering, repairs artifacts by revising the generation prompt. It creates targeted questions about an output, sends them to a vision-language model, and converts the answers into natural-language feedback. The researchers call these instructions semantic gradients because they guide the next generation step in a role similar to numerical gradients during model training.

A global selection mechanism limits overcorrection by retaining every candidate produced during the optimization sequence. A rater compares those candidates with the original prompt and selects the highest-scoring result. Google’s examples include correcting a rigid cuboid so it resembles an inflated balloon and preventing musicians from exchanging instruments between shots.

Benchmarks stress long gaps and changing state

Google introduced three benchmarks alongside the systems and also evaluated them on existing video-generation suites.

The papers report a peak quality score of 81.4 for AI Video Co-Director on GenAD-Bench, along with higher story-consistency results on ViStoryBench. CANVAS reports continuity gains on ST-Bench and HardContinuityBench. A²RD reports stronger character and environment consistency on VBench-Long and LVBench-C, while VQQA reports compositional gains on T2V-CompBench, VBench2, and VBench-I2V.

Google built the three new benchmarks and evaluated the systems that target them, so independent replication remains an open step. The papers provide the metric definitions, baselines, and category-level results needed to interpret the reported gains.

The reusable layer sits above the generator

The common architecture treats continuity as explicit pipeline state. A comparable implementation would need canonical records for characters and assets, retrieval before each generation call, separate controls for narrative progress and continuity, candidate-level evaluation, and final selection against the original brief.

Repeated generation, multimodal judging, and global candidate selection also add inference calls, latency, and storage requirements. Teams adopting this pattern would need to budget for those costs and define stopping rules for iterative refinement. The control methods could wrap other foundation video models, although Google’s published experiments use Gemini and Veo.

Current access stops at papers and benchmarks

Google has published preprints for all four systems. Its announcement identifies COLM 2026 for AI Video Co-Director and EMNLP 2026 for CANVAS. A public API and open-source implementation remain unavailable.

The benchmark repositories are linked from Google’s project materials, allowing developers to test other pipelines against the same continuity scenarios. For now, implementation requires reconstructing the orchestration patterns from the papers and supplying separate planning, generation, memory, and evaluation components.