# HTML is All You Need (for Agents to Make Graphics)

Amol Kapoor, Nori | AI Engineer World's Fair 2026 | 07:00

Source: https://www.youtube.com/watch?v=JRTAtZ5iBkU
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/html-is-all-you-need-for-agents-to-make-graphics
Published: 2026-06-28
Tags: agents, coding-agents, tool-use

## TL;DR
- Coding agents can make visual artifacts when they work in a medium built around language and structure.
- HTML lets agents describe headings, charts, grids, fonts, and motion without placing every pixel by coordinate.
- Nori uses HTML to create slide decks, documents, videos, and end-to-end presentations from company data.

## Summary
Amol Kapoor argues that coding agents are poor at graphics mainly because people give them tools designed for human hands and eyes. Canvas tools such as PowerPoint, Figma, and Canva require clicking, dragging, resizing, and visual correction. SVG is also a poor fit because it presents the model with a wall of numbers. HTML gives the agent a language-based representation of layout. Tags describe headings, charts, grids, and other elements, while the browser turns that structure into pixels. Kapoor says this approach works for slide decks, board decks, sales decks, documents, and videos. The content can come from company data such as call transcripts and emails, allowing an agent to build a deck from source material through presentation. His practical advice is to stop treating PowerPoint as the deck itself. Use HTML as the editable format, then render it to PDF or another format when needed.

## Key ideas
### Coding agents can do more than write code
[00:07](https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=7s)
Amol Kapoor begins by rejecting the narrow description of coding agents as tools that only write code. Nori deploys an AI employee that works with company code, documents, Slack, and other data. Kapoor says the important change in approach is to think like an agent when asking it to do work. The talk applies that idea to visual artifacts, including slides, documents, and video. His claim is that the agent's abilities are broader than the name suggests, but the user has to choose an interface that matches how the model operates.

### Most slide-deck work is formatting work
[00:48](https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=48s)
Kapoor says the world spends around 34,000 human years making slide decks each day. In his view, much of that time goes into fiddling with formatting, branding, and object placement rather than thinking. A deck that takes 10 hours could take about 25 minutes if those tasks were removed. Traditional tools begin with a canvas in PowerPoint, Google Slides, Figma, or Canva. The user clicks, drags, drops, resizes, and snaps objects to a grid. Those actions fit human vision and hands, but they do not give an agent a useful way to express the intended structure.

### Canvas tools give agents the wrong interface
[01:33](https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=93s)
When agents use visual editing tools, Kapoor says objects overlap, text becomes unreadable, and alignment breaks down. Skeptics interpret these failures as evidence that agents cannot reason about space. He mentions ARC-AGI and Simon Willison's pelican-riding-a-bicycle test as examples of that concern. In the test, the model must draw the image using SVG. Kapoor says the poor results are expected because SVG is a wall of numbers. A person would also struggle to handwrite an SVG, since neither people nor models can easily turn raw coordinates into a visible scene.

### HTML matches the model's language-based way of thinking
[03:00](https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=180s)
Kapoor argues that the medium is the problem. Agents work with words, tokens, and structure, so the interface should describe graphics in those terms rather than in pixels. HTML has semantic tags for elements such as headings, charts, and grids. The browser handles the conversion from that structure to pixels. This means the model can describe layout without manually placing every coordinate. Kapoor says HTML also supports visual effects, charts, layouts, fonts, and motion. An HTML version of the pelican task gives the model code that it can read, theme, and edit line by line.

### A slide deck does not need PowerPoint as its editing format
[04:30](https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=270s)
Kapoor separates the presentation from the application used to create it. He spent much of his life making decks in PowerPoint and assumed the two were synonymous. He now treats PowerPoint as one possible editing tool, while the deck is what the audience sees in presentation mode. The editing format is arbitrary because the audience does not care how the final presentation was produced. HTML can therefore be the working format for agents, with the result rendered to PDF or another required format later.

### Nori uses HTML for decks, documents, and video
[05:11](https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=311s)
Kapoor says Nori uses the HTML approach for slide decks, board decks, and sales decks that the company presents and sends regularly. The same method is used for documents, where HTML adds color and follows the company's brand. He also says the video itself is made with HTML and CSS, describing it as 'just divs all the way down.' His broader point is that plain text is often a convenience choice. When the goal is to make something useful for an audience, structure and color can communicate more than unformatted text.

### Agents still need source content and a story
[05:53](https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=353s)
Kapoor qualifies the graphics argument by saying that a beautiful deck alone is generally not valuable. The agent still needs the content that populates it. If the model can access data such as call transcripts or emails, it can build the deck end to end. He describes this as handing the grunt work to agents while a person focuses on vision and story. Nori Sessions supports this workflow, and Kapoor says he has built entire board decks from his phone during his subway commute because the Nori bot is connected to the company's internal information.

### The practical instruction is to choose the model's native medium
[06:33](https://www.youtube.com/watch?v=JRTAtZ5iBkU&t=393s)
Kapoor closes with a direct instruction: stop thinking like a user and think like the model. The user should provide a language the agent already understands well. For visual output, he identifies HTML as that language because it combines readable structure with browser rendering. This avoids asking the model to operate a human-oriented canvas or to reason from raw SVG coordinates. His final claim is deliberately narrow and practical: for agents making graphics, HTML is enough.

## Notable quotes
- "Coding agents can do almost anything." (00:32)
- "If you ask me, it's not the model, it's the medium." (02:38)
- "HTML lets the model think in structure." (03:45)
- "Stop thinking like a user. Think like the model." (06:33)

## Tools & references mentioned
- Nori
- Nori Sessions
- PowerPoint
- Google Slides
- Figma
- Canva
- SVG
- ARC-AGI
- Simon Willison
- HTML
- CSS

## Who should watch
- You are asking an agent to create slides or documents and keep getting broken layouts, overlapping objects, or unreadable text.
- Your team spends more time formatting decks than developing the content and story.
- You want an agent to turn internal material such as emails or call transcripts into a presentation that can later be rendered to PDF.