In April, Andrej Karpathy printed a GitHub gist file known as “LLM Wiki,” a short textual content doc designed to assist one construct a private data base utilizing LLMs. It’s primarily based on the premise that an AI agent will maintain what it is aware of as linked Markdown information it could learn and rewrite, as a result of a language mannequin doesn’t get bored sustaining cross-references and might contact fifteen information in a single move. It was only some thousand phrases with no product connected.
Two months later, Google turned that intuition into a broadcast customary known as the Open Knowledge Format. The OKF packages organizational data, metrics, tables, and runbooks as plain Markdown that any agent can learn and not using a proprietary account. Google is cautious to name it v0.1 — a place to begin slightly than a completed customary.
Garry Tan, the Y Combinator president, acquired there first in a special lane. His gstack, an MIT-licensed Claude Code setup that crossed 66,000 GitHub stars inside weeks, contains 23 specialist roles, every a Markdown file. No runtime; no code; simply prose that runs throughout ten totally different coding brokers.
Markdown has change into the substrate brokers learn and write
Three approaches, three totally different wants, one widespread answer. Karpathy sought agent reminiscence, Google aimed for enterprise context in BigQuery brokers, and Tan wished a option to summon an engineering workforce from a terminal. All three turned to the identical fundamental useful resource: a folder of Markdown information versioned in git.
Builders had already established this follow. CLAUDE.md and AGENTS.md are current in hundreds of thousands of repositories because the preliminary information an agent masses. OKF and gstack are the developed types of this conference – one centered on what the agent is aware of, the opposite on the way it behaves.
That is the Git and JSON playbook tied to the agent’s data. The codecs that survived are those you could possibly begin utilizing with out altering something. You may merely cat the file, clone the repo, and any instrument you already use can parse it. MCP stays essential because the interface an agent connects to. Markdown is turning into the format that carries the content material.
The lock-in moved from the mannequin to the information
The numerous issue to look at right here is the aggressive benefit, not technical specifics. For 2 years, the assumption was that proudly owning the perfect mannequin meant controlling the developer.
This angle is now shifting. Changing Claude with GLM or Codex, gstack continues to function as a result of the core intelligence developed, however the documentation didn’t.
The moat is shifting from the mannequin to the Markdown a workforce owns and accumulates over time.
The moat is shifting from the mannequin to the Markdown a workforce owns and accumulates over time. An organization’s OKF bundle, together with its runbooks, metric definitions, and structure choices, is, by design, transportable throughout clouds, fashions, and frameworks.
That sort of portability is the explanation vendor-neutral codecs exist and why Google’s OKF deserves a better look.
If nobody develops shoppers for it, it stays simply a good suggestion that Google launched on a sluggish Friday.
The world the place I’m almost definitely mistaken is sturdiness. Declaring Markdown requirements is straightforward, however making them dependable is tough. OKF is merely a 0.1 draft with a reference implementation, not a full ecosystem. If nobody develops shoppers for it, it stays simply a good suggestion that Google launched on a sluggish Friday.
The path stays decided by three separate bets focusing on the identical file format inside a single quarter. Your subsequent agent is more likely to interpret its context from a Markdown folder, and the creator of that folder now possesses a bonus that the mannequin vendor can’t simply replicate.
YOUTUBE.COM/THENEWSTACK
Tech moves fast, don’t miss an episode. Subscribe to our YouTube
channel to stream all our podcasts, interviews, demos, and more.
