I am Jarvis Nelson Osorio, also known as Jarvis Osorio, a technology entrepreneur and innovator based in Chile. Before I talk about agents, I want to talk about a quieter standard: plain text with just enough structure to be kind to humans and to machines. That standard is Markdown, and it is having a second life in AI systems.
I originally shared this history and argument on LinkedIn in The Plain-Text Revolution: A Brief History of Markdown and Why It Powers Modern Software and AI.
A 2004 writing hack that refused to die
In the early web you had an ugly choice. Raw HTML was heavy with tags. Rich editors locked writing into brittle, often binary formats that fought version control and rendered differently everywhere. Plain .txt files traveled well but had no hierarchy—no honest heading, list, or link.
In 2004, John Gruber, working with Aaron Swartz, proposed Markdown: a lightweight markup you can read as ordinary text, with simple markers for headings, emphasis and lists, that can be turned cleanly into HTML. It was a writing tool. Then it became the language of software documentation, because Git can show a useful diff of a .md file the way it cannot show a useful diff of a sealed document.
Why language models like the same files
Agents and retrieval pipelines thrive on text they can parse without a ceremony. Markdown is compact, so you spend fewer tokens on decoration. It has a visible skeleton, so a model can find the section that matters. It can live next to code in the same repository, which means your product facts and your product source can share a history.
When teams keep the source of truth in slide decks and locked files, every agent has to guess. When the source of truth is a versioned Markdown corpus—product requirements, runbooks, decision records—the agent has a map. That is not a small convenience. It is how you ground a system in something other than a confident improvisation.
How I treat it in an agentic workflow
I want a living map between research, design and code, and I often want that map to be a plain file the whole team can read. The same instinct shows up in the "design.md" bridge I have described in other writing: give the agent a document it can revisit, not only a chat it will forget. Persistence favors formats that survive a handoff.
This is also a cultural choice. Fancy tools come and go. A folder of Markdown, in a repository your team actually owns, is a kind of institutional memory. It is closer to a constitution than to a mood. That is why it sits comfortably beside the arguments in Breaking the Centralized Order: if the rules of a system cannot be read, they cannot be contested.
If you are building with language models this year, do not start with a new dashboard. Start by putting the sentences you already trust into files a human can read, a git history can track, and an agent can retrieve without getting lost in formatting.
Where to go next
This expands on ideas I originally shared on LinkedIn in The Plain-Text Revolution: A Brief History of Markdown and Why It Powers Modern Software and AI.
More essays, books and profiles: jarvisosorio.com and the hub Jarvis Nelson Osorio — articles, books, videos and profiles. For how institutions need readable rules, see Breaking the Centralized Order.
More from Jarvis Osorio: https://jarvisosorio.com
Top comments (0)