Descriptions:
Amol Kapoor, CEO of Nori Agentic, delivers a concise AI Engineer conference talk arguing that HTML — not SVG, not PowerPoint CLIs, not canvas-manipulation loops — is the correct native medium for AI agents producing visual artifacts like slide decks, documents, and video.
The argument starts from first principles: existing tools for making slides (PowerPoint, Figma, Canva) are designed around human spatial reasoning — click, drag, snap to grid. When agents use them, output is misaligned and unreadable. The popular Simon Willison “pelican on a bicycle” SVG test illustrates why: SVG is a wall of coordinates that neither humans nor models can reason about intuitively. Kapoor’s thesis is that the failure is not the model, it is the medium. HTML, by contrast, is a structural language that models have trained on billions of examples of. It maps directly to how language models represent information — as tokens and hierarchy, not pixels and coordinates. The browser handles the pixel rendering for free.
Nori uses this approach in production: their board decks, sales decks, documentation, and even the video of the talk itself are generated as HTML and CSS. Kapoor demonstrates that the same agent capability — coding — that most people think of narrowly can produce richly formatted, on-brand visual output when given the right output format. The talk closes by noting that beautiful output is still empty without substantive content, pivoting to how agent-native knowledge retrieval fills that gap.
📺 Source: AI Engineer · Published June 28, 2026
🏷️ Format: Keynote Launch
