Formats for Agents: Why HTML, JSON and Markdown Win

A Word document or a PowerPoint slide is a mix of content, formatting and layout. When an AI agent is asked to change a number in it, it risks shifting the layout. That is not a flaw of the AI but of the format's structure. Internally, these files are zipped archives of nested XML in which a single sentence can be spread across several formatting runs.

Separating content and presentation

In a structured format, content is stored as data: headings, fields, values. Presentation comes from a separate template. The agent changes exactly the field that needs changing, and the presentation stays untouched. HTML, JSON and Markdown are the common formats for this: human-readable, versionable and processable by any tool.

Each format has its role. JSON carries structured data such as line items, amounts and dates. Markdown carries running text with simple structure. HTML combines both with a presentation that works in any browser. Because all three are plain text, every change can be compared line by line: who changed which value and when is visible at a glance.

Getting to the familiar file

Clients often expect Word or PowerPoint, and that remains possible. The file is generated at the end of the process from the structured data, not at its beginning. You keep the familiar format and gain the benefits of a clean source: fast changes, consistent layouts and text that an agent can edit reliably.

At its core, this is about a shared language between people and machines. We described that idea in our article on no-code and AI. Structured formats are the practical step towards it.

Three takeaways

1. Structure beats layout. What exists as data can be changed, checked and reused precisely.
2. The file is a result, not a starting point. Word and PowerPoint remain as output; the source becomes open and structured.
3. Open formats are future-proof. They work with today's and tomorrow's tools, without vendor lock-in.