Three people, one format, three different problems
The observation that makes the article worth keeping: three of the most-watched people in AI reached for the exact same primitive within one quarter — plain .md files in a git repo — while trying to solve completely different problems.
Karpathy — agent memory
In April he published a GitHub gist called “LLM Wiki” — a few thousand words, no product attached. The premise: an agent keeps what it knows as linked Markdown files it can read and rewrite, because — in his words — a language model “does not get bored maintaining cross-references and can touch fifteen files in a single pass.”
Google — enterprise context
Two months later Google published the Open Knowledge Format (OKF): organizational knowledge, metrics, tables, and runbooks packaged as plain Markdown that any agent can read without a proprietary account. Aimed at BigQuery agents. Notably shipped as v0.1 — a starting point, not a finished standard.
Garry Tan — a team from a terminal
The YC president’s gstack — an MIT-licensed Claude Code setup that crossed 66,000 GitHub stars in weeks — is 23 specialist roles, each a Markdown file. No runtime, no code: “just prose that runs across ten different coding agents.”
Karpathy wanted agent memory. Google wanted enterprise context. Tan wanted to summon an engineering team. Same resource underneath all three: a folder of Markdown files in git.
| Who | Artifact | Problem they were solving |
|---|---|---|
| Karpathy | “LLM Wiki” gist | Agent memory the model maintains itself |
| Open Knowledge Format (v0.1) | Portable enterprise context for agents | |
| Garry Tan | gstack (23 role files) | A specialist engineering team, as prose |
Why Markdown, and why now
Developers had already set the convention. CLAUDE.md and AGENTS.md sit in millions of repositories as the first files an agent loads. OKF and gstack are just the evolved forms of that habit — one focused on what the agent knows, the other on how it behaves.
“The formats that survived are the ones you could start using without changing anything. You can simply
— Janakiram MSV, The New Stackcatthe file, clone the repo, and any tool you already use can parse it.”
The framing that stuck with us: this is the git-and-plaintext playbook applied to an agent’s knowledge. MCP remains the interface an agent connects through; Markdown is becoming the format that carries the content across. Two different layers, and it’s worth not confusing them.
The actual claim: the moat is moving
“The moat is shifting from the model to the Markdown a team owns and accumulates over time.”
— The New Stack
This is the line worth arguing about. A company’s knowledge bundle — its runbooks, metric definitions, architecture decisions — is, by design, portable across clouds, models, and frameworks. The model is rented and swappable. The accumulated context is owned and compounding. If that’s right, the defensible asset is not which frontier model you called this week; it’s the corpus your agents read from — and the person who authored that corpus “now possesses an advantage that the model vendor cannot easily replicate.”
Janakiram flags where he might be wrong: durability. “Declaring Markdown standards is easy, but making them reliable is difficult.” OKF is a 0.1 draft with a reference implementation, not an ecosystem — if nobody builds consumers for it, it stays “a good idea Google released on a slow Friday.” A format is only a moat once things reliably read and write it.
Why this rhymes with two of our other notes
We’ve now written the same underlying idea from three directions, and they reinforce each other:
- In the Asana note, the compounding asset is the work graph — every AI approval becomes training data the incumbent owns.
- In the Iron Man Principle note, the leverage comes from write-back — agents that record what they do, not just chat.
- Here, the asset is the Markdown corpus — portable, model-agnostic, owned.
Three vocabularies, one thesis: the durable value in AI is accumulated, structured context — not the model in front of it.
What this means for VeehiveLabs
If the moat is the knowledge a team owns, then the most valuable thing we can leave behind on an engagement isn’t only a working integration — it’s the portable knowledge layer that makes the customer’s agents smart, and keeps working no matter which model or vendor they choose next. That’s the same discipline as our forward-deployed rule (“convert custom work into reusable modules”), pointed at the customer’s side of the fence.
On every AI build, treat the customer’s AGENTS.md / OKF-style knowledge files as a first-class deliverable: runbooks, metric definitions, glossary, decision records — captured as plain Markdown in their repo. The integration wires things up; the knowledge layer is what compounds.
Keep the knowledge model- and vendor-neutral. If a client’s context only works inside one proprietary agent, we’ve built them a dependency, not an asset. Plain Markdown in git means they can switch models, clouds, or frameworks and carry their advantage with them — which is exactly why they’ll trust us to build it.
The person who authors the folder holds the advantage. We should be excellent at authoring and maintaining these knowledge files — a repeatable Veehive practice for structuring an organisation’s context so agents can actually use it. That authoring skill is itself a service we can sell.
What we take away
- Convergence is a signal. Three independent, credible bets on the same primitive in one quarter is rarely coincidence — Markdown-as-agent-memory is a real pattern, not a fad.
- Separate the layers. MCP is the interface; Markdown is the content. Don’t let one absorb the other in how we design systems.
- The moat is the corpus, not the model. Accumulated, portable context is the compounding asset. Models are swappable; owned knowledge isn’t.
- Durability is the open risk. A format is only a moat once tools reliably read and write it. Watch whether OKF grows real consumers before betting a client’s architecture on it.
- Our deliverable includes the knowledge layer. Build and hand over the portable Markdown corpus — that’s where the client’s durable advantage (and our differentiation) lives.
Source: The New Stack — Janakiram MSV, “Andrej Karpathy, Google and Garry Tan agree Markdown is the answer, but they’re not solving the same problem” (6 July 2026). Quotations are from the article; emphasis and the VeehiveLabs section are ours. Notes compiled by the VeehiveLabs team; corrections welcome.