Agentic Workflow for 1-Hour Movie Generation

15-minute screen for the Video-Generation / Agent-RL team. After a quick ML-direction intro, the interviewer asks: how would you design an agentic workflow that generates a one-hour-long movie? Open-ended, no right answer — the bar is hypothesis quality and how concretely you can scope subproblems under time pressure. Requirements In ~10 minutes verbal, sketch: The high-level pipeline (script → shots → frames → audio → assembly) and the agents that own each stage. The consistency problem: characters, environments, and props must remain coherent across thousands of generated frames spanning many scenes. Narrative pacing: the agent must decide scene length, transitions, and tonal arcs — not just generate frames. The memory / context layer the agents share: scene bible, character sheets, continuity tracker. Where reinforcement learning fits (reward model on coherence + viewer engagement, RLAIF for narrative quality). Compute and latency tradeoffs: per-frame inference cost × frames × scenes × revisions. Notes The interviewer probes for concrete subproblems rather than a wishlist. Pick two or three to go deep on (consistency tracking and narrative pacing are the most defensible choices) instead of skimming the whole pipeline. A reasonable hook: a "director" agent decomposes a script into shot list; "consistency" agent maintains an embedding-keyed memory of characters / props; "renderer" agent calls the underlying video diffusion model with conditioning from both; "editor" agent does cuts and pacing review. The reporter felt blindsided in the moment — the round rewards candidates who have thought about agent decomposition for long-horizon generation tasks beforehand. Mention real systems if you can (Sora, Runway Gen-3, Veo) but treat them as inspiration rather than reference architectures — the interviewer wants your take, not a recap. Preparation Read one survey on long-horizon video generation and one on multi-agent LLM workflows; you only need vocabulary and intuition. Prepare a 90-second pipeline pitch you can deliver from memory; spend the rest of the round on whichever subproblem the interviewer picks. Practice one ML-direction intro that can pivot into video / multimodal — this round assumes you bring some background. Concrete production constraint to ground the design: current state-of-the-art end-to-end video diffusion (transformer-backbone, recaptioned text-conditioning, multi-frame foresight for subject persistence) tops out around 20 seconds at 1080p per single generation call. A one-hour movie therefore decomposes into ~180 clip-generations that must be stitched coherently, so the agentic layer above the model is doing the heavy lifting — the diffusion model itself is one tool the director-agent calls many times, not the answer. Two consistency layers worth naming explicitly: (a) intra-clip — handled by the underlying diffusion model's multi-frame foresight ("keep a subject the same even when it leaves frame"), so do not re-derive it; (b) inter-clip — handled by an external memory of character embeddings / scene bibles / continuity tracker, which is the candidate's design contribution. A defensible reward signal for the RL layer: per-clip CLIP-style alignment with the script + a coherence model that scores cross-clip continuity (same character embedding, consistent lighting); aggregate into an editor-agent value function.