Better video generation does not automatically produce a coherent film or a reliable interactive world. The key shift examined here is from treating the generator as the whole system to treating it as one component in a harness that plans, acts, tracks state, and learns from feedback.
Key Findings
- Production is a coordination problem: Script and shot planning, asset management, generation, and editing must share project context. This makes consistency and workflow quality separate evaluation targets from the visual quality of individual clips.
- World state need not live inside the video model: A game engine or agent-written program can maintain state while a video model renders observations. This hybrid design can support more controlled interaction, though it raises questions about how faithfully generated frames reflect the underlying state.
- Improvement has several distinct targets: Revising a clip is not the same as changing harness memory, generator parameters, an evaluator, or the procedure that selects future updates. Claims of recursive self-improvement require retained changes and independent evaluation across later rounds.
- Evaluation must connect quality to sustained progress: The report highlights provenance, cost accounting, control fidelity, and long-horizon consistency as necessary evidence—not just attractive outputs or a single successful task.
Who Should Read It
Great fit for researchers designing agentic video systems, video world models, or model-development loops who need a framework for separating system components and evaluation claims. Look elsewhere for a new video-generation architecture, benchmark results, or a ready-to-run implementation: this is a conceptual survey and research agenda, and many proposed mechanisms remain open challenges.