From prompt to playable: media agents learn to run the production

The whole editing room in one agent

WeVisCut, in a preprint, is a script-to-video system that trains a compact policy to coordinate video editing tools. It achieves a 90.5 overall score on a 100-task benchmark, and uses 33 times fewer generated tokens per successful output minute than Claude Opus 5. The token number matters more than the benchmark score. Video is long, and a system that burns tokens on every minute of output cannot be run at any serious volume, however good a single clip looks. m:…4454

The preprint also describes how the system was built to be cheap to run: its data pipeline yielded 134,771 training examples from 3,750 finished videos, combining editing inversion with verified replay, and the system uses a single 9B policy to coordinate production and recover from failures. The design bet is that editing is a coordination problem. One small policy directing the tools, with a way to recover when a step fails, beats a large model improvising each cut from scratch. The preprint reports that WeVisCut outperforms task-matched native editors on its proposed metrics, though those metrics are the authors' own. m:…4454

Games need a state, not just a scene

VibeGame, in a preprint, applies the same coordination idea to a harder target: building 2D games from natural-language prompts. Its engine represents project state as schema-validated structured text and provides frame-synchronous runtime control, and it uses an Adversarial Agent Team of eight specialized agents that separates implementation from review. The separation is the interesting choice. A game is not a picture; it has rules, and rules break in ways that only an adversary looking for the break will find. m:…6838

The system also includes a training-free self-evolution mechanism that distills projects into reusable skeletons and collaboration contracts, and it supports both prompt-to-game and prompt-assets-to-game tasks, accepting optional reference images in prompts. On GameCraft-Bench, the preprint reports that VibeGame outperformed the strongest baseline by 11.70 points using GPT-5.6-Luna and by 4.19 points using GPT-6.1-Sol. The gap between those two margins is itself informative: the harness helps, but how much it helps depends on the model underneath it. m:…6838

Priced by the second, automated by the brief

Kling 4.0 shows what this looks like as a commercial product rather than a paper. The company says it generates 3 to 30 seconds of 720p or 1080p video with two-channel stereo audio, and that it generates its own two-channel stereo soundtrack timed to the video. On fal, the company prices generation at $0.157 per second at 720p and $0.235 at 1080p, with video input billed separately, and generation billed on input plus output seconds and never less than twice the output. A 30-second clip is therefore a line item, not a toy, and anyone budgeting a campaign should do the multiplication before falling in love with a demo. m:…8074

The company also says fal Agent can automate multi-step Kling 4.0 generation workflows from a single brief, that fal offers four endpoints covering text-to-video, image-to-video, reference-to-video, and video-to-video or edit, and that reference-to-video endpoints accept up to 15 reference assets per request, with product consistency maintained by creating elements from one frontal still and up to three extra angles. That is the same pattern as the research papers above, sold as infrastructure: the brief goes in, an agent runs the steps, and consistency across shots is treated as a feature to engineer rather than luck to hope for. m:…8074

A world that remembers without ballooning

L3G-World, in a preprint, tackles the memory problem that long-form generation runs into. It uses a compact latent 3D Gaussian memory and alternating autoregressive updates to generate spatially consistent interactive worlds, and its memory grows with scene extent rather than stream length. Existing 3D memory approaches scale poorly, the paper notes, because memory grows with the length of the generated stream. A world that gets more expensive to remember the longer you explore it cannot stay interactive for long. m:…1750

The preprint reports that L3G-World achieves real-time world generation at 9 FPS, uses up to 38.6 times fewer 3D primitives than the most efficient prior 3D-memory baseline, and is 3.2 times faster in memory updates and 359 times faster in memory rendering than that baseline. Those are the authors' comparisons, not independent measurements, but the direction matches the rest of the week's evidence. Across video, games, and worlds, the competitive edge is moving from generating a striking first frame to holding a coherent, affordable production together over time. m:…1750