Video Generation Meets Agent Harness: Agentic Video Generation, Agentic World Modeling and Recursive Self-Improvement

The paper frames video generation through an agent harness that coordinates production, world-model interaction, and development toward recursive self-improvement.

Video generators can produce increasingly realistic and controllable clips, but a generator alone does not organize the decisions needed to complete a video task: a brief must become coherent shots, references must persist across scenes, takes must be inspected, and edits must preserve what is correct. Interactive video must also support continuing actions grounded in observations, not just a finished clip. The paper notes that plausible footage may still violate a narrative requirement or fail to realize an intended action, so a quality score alone rarely specifies what to do next.

The paper is a survey/report organized around the agent harness as the unit of analysis across production, interaction, and model development. It defines a common agent-harness execution loop and applies it to agentic video generation, organized by workflow function, and to world-model interaction, organized by the interfaces through which a harness intervenes. It examines designs in which an engine or a program written by a coding agent holds the world state while the video model renders it. It separates revisions of a current output from updates to harness state, generator parameters, and evaluators, and from changes to the procedure that produces later updates. It identifies four channels through which an improvement could re-enter video development and concludes with evaluation requirements and open challenges.

• The report is organized around three directions: agentic video generation, agentic world modeling, and recursive self-improvement. §Abstract
• The paper's contributions are threefold: a harness-centered framework, an account of improvement, and evaluation requirements and open problems. §1.2
• The paper separates revisions of the current output from updates to harness state, generator parameters, and evaluators, and from changes to the development procedure. §1.2
• The paper identifies four channels through which an improvement could re-enter video development. §1.2
• No surveyed system yet revises the procedure that trains and validates the generator. §1.2

• The paper states that no surveyed system yet revises the procedure that trains and validates the generator.
• The paper leaves open whether the mechanisms that construct and update reusable knowledge can themselves improve.
• The paper states that a harness revision must be retained and independently evaluated on later tasks or improvement rounds under stated resource constraints; this does not make every harness self-improving.

• The agent harness is the executable runtime around the model that constructs context, exposes tools, maintains state, routes feedback, and enforces permissions and budgets. §1.1
• Agentic video generation is a system-level perspective rather than a new generator architecture; the agent makes consequential choices around synthesis or predicted transitions. §1.1
• Feedback changes different objects at different timescales, including the current artifact, subsequent tasks, and later model versions. §1.1
• The survey connects three questions: harness organization for production, harness support for world-model interaction, and movement toward RSI. §1.2
• In agentic world modeling, harnesses mediate observation, control, execution, and state, including designs where an engine or coding-agent program holds world state while the video model renders it. §Abstract
• The paper separates current-output revision from updates to harness state, generator parameters, evaluators, and the procedure producing later updates. §1.2
• The paper lists open challenges including autonomous video production, lifelong memory, programmable hybrid control, autonomous model development, provenance and cost accounting, and evaluation linking video quality and task performance to sustained improvement. §Abstract
• The paper differentiates itself by taking the harness around a video generator as the unit of analysis across production, interaction, and development. §1.2