AGENTIC VIDEOGEN
Preprint 2026

Video GenerationMeetsAgent Harness.

  • Video Generation
  • Agent
  • Harness Engineering
  • World Models
  • Recursive Self-Improvement
INSIDE THE STUDIO
A little robot making a video in its studio The white and blue Agent Harness robot from Figure 1 sits at its computer and types on the keyboard. It has round blue ears, a ball antenna, a dark face with smiling cyan eyes, and a five-module tool belt. A landscape takes shape on the monitor while the editing timeline plays in a continuous loop. VIDEO STUDIO

A little agent. A lot of imagination.

Plan the shot. Generate the scene. Refine the cut.

—PAPERS & RESOURCES
03CORE READING PATHS
—CODE LINKS
Discover the landscape

THE SHARED PERSPECTIVE

One generative foundation.
Three ways to build around it.

The video model is the common ground.
The agent harness connects goals, tools,
memory, feedback, and execution.

01 / CREATE

Agentic Video
Generation

Plan the story. Coordinate assets and tools. Inspect and refine the final cut.

Brief → workflow → video—
02 / INTERACT

Agentic World
Modeling

Act in generated worlds. Maintain state and memory. Test what happens next.

Action → world → observation—
03 / IMPROVE

RSI in
Video Generation

Improve data, models, evaluation, and the procedures that produce later updates.

Experience → update → capability—

Complementary research settings, not a hierarchy of capability levels.

FIGURE 2 · FROM THE PAPER

Explore the research tree.

Select any work to open its paper or project.

Open full figure
Figure 2 from the paper: a shared Video Generation trunk branches into Agentic World Modeling, Agentic Video Generation, and RSI in Video Generation. Four regions indicate increasing agent involvement. Stable Video Diffusion VideoCrafter2 Sora Open-Sora 1.0 CogVideoX Movie Gen Open-Sora Plan HunyuanVideo Dual-IPO Open-Sora 2.0 Wan 2.1 Seedance 1.0 Wan 2.2 GAIA-1 WorldDreamer Genie IRASim AVID GameGen-X Cosmos AdaWorld WorldScore WorldMem WorldExplorer Spatial Memory Matrix-Game Matrix-Game 2.0 WorldPack WorldPlay Matrix-Game 3.0 Mirage VideoDirectorGPT GPT4Motion VideoStudio LAVE Mora MM-StoryAgent MovieAgent Camera Artist AgentMV CutVerse UniPi RoboDreamer GameNGen DINO-WM NWM VideoWorld V-JEPA 2 CoLA-World WorldArena DreamZero FutureVLA WorldArena 2.0 PlayWorld Kubrick DreamFactory StoryAgent GenMAC FilMaster MAViS EditDuet VISTA ELLMPEG CutClaw CineAgents ViMax GroundShot FRAMEWORKERS UniSim DIAMOND iVideoGPT World4RL Dreamer 4 ABot-World-0 VC-Agent SPIRAL JAVEdit AesopAgent VideoAgent AESR SIMA 2 PROWL SILVR VDC-Agent VLAW RoboEvolve VideoWeaver AVA-Encoder CineForge MoCoGAN VideoGPT Video Diffusion Models CogVideo Make-A-Video Imagen Video Phenaki

Branches show thematic connections; vertical position reflects agent involvement.

96 linked works

TABLE 1 · POSITIONING THE SURVEY

Comparison with related literature.

52 related studies · 6 comparison dimensions
Grouped and ordered as in the paper.

Systematic coveragePartial coverageOutside scope
Scroll to explore the full table ↕ ↔
Table 1. Comparison with related literature. 52 related studies and our survey.
Survey / StudyYearScopeThematic CoverageAgentic
Design
Video GenerationWorld ModelsModel DevelopmentAgentic Video GenerationAgent–World Interaction
Visual Generative Modeling
Adversarial Text-to-Image Synthesis2021
Diffusion Models: Methods and Applications2022
RenAIssance: Text-to-Image Generation2023
Text-to-Image Diffusion Models2023
From Sora: Text-to-Video Generation2024
Autoregressive Models in Vision2024
Video Diffusion Models2024
Visual Editing and Control
Diffusion-Based Image Editing2024
Controllable Text-to-Image Generation2024
Diffusion-Based Video Editing2024
Long Video Generation and Memory
Long Video Generation2024
Long Video Generation: Latest Trends2024
Spatiotemporal Consistency2025
The Past Frames the Future: Video Memory2026
Specialized Video Generation
Human Video Generation2024
Controllable Video Generation2025
Physics Cognition in Video Generation2025
Generation Evaluation and Alignment
Diffusion Model Alignment2024
AI-Generated Video Evaluation2024
Preference Alignment for Image Generation2025
Video Post-Training and Alignment2026
Visual Production Workflows
AI-Empowered Persuasive Video Generation2021
LLMs Meet Multimodal Generation and Editing2024
Visual Generation in the New Era2026
Agentic Generative Systems2026
Agentic Artifact Creation2026
Beyond Visual Generation: Agentic Creation2026
Agentic Visual Generation: Models to Control2026
Agentic Visual Generation2026
Video Understanding Agents
Agentic Video Understanding2026
Video-RSI: Harness Evolution (method)2026
General and Interactive World Models
Sora as a World Model?2024
Understanding World or Predicting Future?2024
Interactive Generative Video2025
Visual World Simulation: A Roadmap2025
Agentic World Modeling2026
Quo Vadis, World Modeling?2026
Embodied World Models
World Models for Embodied AI2025
A Step Toward World Models: Manipulation2025
Robotic Video World Models2026
World Models for Robot Learning2026
World Models for Robotic Manipulation2026
World-Action Models for Robot Learning2026
Embodied World Models: Plausible to Actionable2026
Agent Harness
Code as Agent Harness2026
Agent System and Harness Design2026
Self-Evolving Systems
Self-Evolution of Large Language Models2024
Self-Evolving Agents2025
RSI: Bounded Refinement to Research Loops2026
Recursive Self-Improvement in AI2026
Self-Improvements in Modern Agentic Systems2026
The Last AI Built by Humans: Genuine RSI2026
Ours2026

Agentic Design covers agent-based methods; thematic coverage includes enabling methods. Years indicate first release (Ours: manuscript year), and coverage follows the cited version. Titles are abbreviated. Video-RSI is a method paper, so its symbols indicate method scope.

THE OPEN READING ROOM

Find your next paper.

Catalogue loading…

Loading the catalogue…

PRIMARY SOURCES · DIRECT LINKS

A paper can belong to several reading paths. Bibliography metadata is retained where a primary-source refresh is pending; corrections and official code links are welcome.

A COLLABORATIVE RESEARCH SURVEY2026 EDITION
  • Liaoyuan Fan*
  • Yuhao Zheng*
  • Dongxing Mao*
  • Hangyu Ran*
  • Siyuan Hu*
  • Weijia Mao
  • Rui Zhao
  • Weijia Wu
  • Guian Fang
  • Yanzhe Chen
  • Zeyu Zhu
  • Yuxin Jiang
  • Yiren Song
  • Yuchao Gu
  • Qi Hu
  • Dan Kong
  • Alex Jinpeng Wang
  • Kevin Qinghong Lin✉
  • Mike Zheng Shou✉
  • Wei Liu✉

*Equal contribution✉Corresponding author

CITE THIS WORK

BibTeX

If you find our survey useful, please cite:

Download .bib 
@misc{fan2026videogenerationmeetsagent,
  title={Video Generation Meets Agent Harness: Agentic Video Generation, Agentic World Modeling and Recursive Self-Improvement},
  author={Liaoyuan Fan and Yuhao Zheng and Dongxing Mao and Hangyu Ran and Siyuan Hu and Weijia Mao and Rui Zhao and Weijia Wu and Guian Fang and Yanzhe Chen and Zeyu Zhu and Yuxin Jiang and Yiren Song and Yuchao Gu and Junchao Yi and Qi Hu and Dan Kong and Alex Jinpeng Wang and Kevin Qinghong Lin and Mike Zheng Shou and Wei Liu},
  year={2026},
  eprint={2610.40191},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.40191},
}
arXiv:2610.40191

BUILT TO KEEP GROWING

Keep the map moving.

Have a paper to add, an official code link,
or a bibliographic correction? Contribute to the collection.

Suggest a resource
Video RebirthAGENTIC VIDEOGENINSIDE THE STUDIO

A LITTLE ROBOT. A LOT OF IMAGINATION.

Every video starts
with an idea.

A short animated illustration of a small robot planning, generating and editing a video in its studio. Skip the opening at any time, or press Escape.

Planning the next shot…

From an idea to the final frame.
  1. 01 Plan
  2. 02 Generate
  3. 03 Edit