Video Generation Meets Agent Harness: Agentic Video Generation, Agentic World Modeling and Recursive Self-Improvement
Video generation is moving to agent harnesses that plan, call tools, and route feedback. The report maps that field, world models, and self-improvement.
The claim
Agent harnesses now coordinate video models the way they coordinate tools.
The problem
End-to-end video models produce clips with no plan and no feedback.
Multi-shot video needs scripting, shot order, assets, and editing.
How it works
The harness builds context and exposes tools around a video model.
It keeps state, routes feedback, and manages the compute budget.
Agentic video generation coordinates scripting, shot planning, production, and editing.
Hybrid worlds keep state in game engines while video models render.
Recursive self-improvement revises outputs, harness state, parameters, and evaluators.
Figures
What changes
The report names the agent capabilities these systems need.
It lists open problems in provenance, cost accounting, and evaluation.
Limits
This work is a report, not a benchmark or a system.
It gives no measured results for any single design.
Source · arXiv