Recent work shows a shift in how computers create videos. Instead of generating a whole clip in one go, researchers are using agents that plan, pick tools, and refine their output step by step. The biggest advance is the use of a runtime called an agent harness. That harness builds the agent’s context, gives it access to tools, stores its memory, and keeps the budget for using heavy video models under control.

The harness coordinates many tasks: it scripts the plot, decides how many shots are needed, manages assets, and even handles editing and watermarking. It also routes any feedback back to the agent so the agent can adjust its plan. The result is more coherent, controllable, and extendable videos.

At the same time, video world models are emerging. These models produce action‑conditioned observations and predicted futures, meaning that an agent can interact with a simulated environment and anticipate how the scene will change after each action. The harness can act as a mediator between the agent and the world model, allowing the agent to observe, control, and refine its own state.

The survey highlights four ways that improvements can re‑enter the development loop: updating the output, the harness state, the generator’s parameters, or the improvement procedure itself. However, none of the current systems have yet shown a revision of the training and validation procedures for the video generators.

The shift to agentic generation promises videos that can adapt to user input, execute longer narratives, and improve over time. Still, challenges remain such as autonomous production, lifelong memory, programmable hybrid control, and linking video quality to task performance. The field is still early, but the framework of an agent harness and searchable world models lays the groundwork for the next generation of video creation.