Why it matters
James Russo describes HeyGen’s path from large prompts and agentic retries to an HTML, CSS, and JavaScript authoring layer for deterministic video. React-based Remotion constrained generated output, while plain web primitives restored flexibility; the team then built rendering controls, evaluation loops, and reusable skills so smaller models could produce consistent MP4s.
My takeaway: Code-native media gives agents an editable artifact and a render-inspect-revise loop, but determinism and safety come from the surrounding runtime. Pin assets and fonts, sandbox generated code, block unexpected network access, test repeatable rendering, and evaluate visual quality across representative templates instead of relying on a polished demo.