Agentic AI
Markdown Files Become the Real Competitive Edge as AI Agents Move Beyond Model Lock-in
Three major figures have converged on Markdown as the foundation for agent memory and knowledge, signaling a shift where content ownership matters more than model choice.
The New Stack
A quiet consensus is forming: Markdown files, not language models, are becoming the true competitive moat in AI development. Within weeks this summer, three separate efforts—a framework for agent memory, an enterprise data standard, and a multi-agent engineering toolkit—all landed on the same answer: plain-text Markdown stored in version control.
Why three different leaders chose the same format#
Andrej Karpathy's approach centered on building personal knowledge bases for LLMs, premised on the idea that agents can maintain and cross-reference sets of interconnected Markdown documents without fatigue—tasks that would exhaust human curators. His GitHub-hosted "LLM Wiki" laid out this pattern in April.
Google's response was more formal. The company published the Open Knowledge Format, positioning Markdown as a vessel for organizational runbooks, metrics, and architecture decisions—content designed to flow seamlessly between BigQuery agents and other tools without requiring proprietary accounts or vendor lock-in.
Garry Tan's entry took yet another path. His gstack project, which accumulated 66,000 GitHub stars in weeks, treats Markdown files as modular job descriptions—23 specialized roles that orchestrate Claude Code and other coding agents without runtime or compiled code, just prose.
Markdown as the common substrate#
Each effort solved a distinct problem, but all converged on the same medium: a folder of versioned Markdown files. Developers had been priming this ground already; templates like CLAUDE.md and AGENTS.md were standard fixtures in millions of repositories long before any formal standard existed. These new frameworks simply formalized and extended what was already emerging in practice.
The appeal lies in simplicity. A Markdown file requires no proprietary parser, no special infrastructure—you can cat it, clone the repo, and feed it to any tool already in your workflow. Unlike JSON schemas or domain-specific languages, Markdown files proved immediately usable without migration costs.
