Google's OKF — Structuring Your Knowledge for Agents

ELI5/TLDR

Google proposed a new file format called OKF (Open Knowledge Format) for packaging up what you know so AI agents can use it. Underneath, it’s nothing fancy — just a folder of plain markdown files, one per concept, with a few standard fields at the top. The pitch is that because the format is standardized, your agent can read my knowledge without needing my software, and experts could eventually sell “knowledge bundles” the way they sell consulting today. The presenter is genuinely excited; whether any of this catches on is still an open question.

The Full Story

What OKF actually is

Strip away the excitement and OKF is a folder of markdown files. Markdown is just plain text with a few marks for headings and bold — the same thing this very note is written in. Each file holds one concept: not a copy of a web page, but a single idea pulled out of it. A business website might decompose into 10 or 50 of these concept files.

Every file opens with a small block of structured fields — type, title, description, tags, timestamp — called YAML front matter. (If you’ve ever looked at the top of an Obsidian note, you’ve seen it.) Below that block, you write whatever you want: instructions, data tables, prose. The folder also carries an index file and a log file, where an agent can record “I updated this today.”

OKF is an open human and agent-friendly format… for representing knowledge, the metadata, the context, and the curated insight that surrounds data and systems. It’s designed to be authored by people, generated by agents, exchanged across organizations, and consumed by both.

The whole collection is called a knowledge bundle. Files can link to each other, so the bundle slowly becomes a web of connected ideas — a knowledge graph. Think of it like a personal Wikipedia for one business, where every article knows which other articles it relates to.

Why “standardized” is the load-bearing word

None of the pieces are new. People have stored markdown in GitHub or Obsidian for years. The claim Haynes keeps returning to is that OKF fixes the format, so any agent can read any public bundle without bespoke plumbing. No special reader, no custom integration — point an agent at a bundle and it understands the structure.

She frames this as a shift away from SEO. The old game was getting found by search. OKF, she argues, is about making a business usable by agents — letting them actually do things with your knowledge rather than just rank a link.

The Karpathy connection — a wiki that writes itself

The more interesting half borrows from Andrej Karpathy’s “LLM Wiki” idea. The usual way to feed documents to an AI is called RAG — retrieval-augmented generation — where you dump a pile of text on the model and make it hunt for the relevant bits at question time. The model does the heavy lifting every single time.

Karpathy’s twist: instead of searching raw documents on demand, the model maintains a living wiki.

Instead of just retrieving from raw documents at query time, the language model incrementally builds and maintains a persistent Wiki.

When you add a new source, the model doesn’t just file it away. It reads it, extracts what matters, and folds it into the existing wiki — updating the relevant pages, revising summaries, and flagging where new information contradicts old claims. So if this video entered Haynes’s OKF today, “OKF” wouldn’t become a brand-new entry — the model would notice the concept already exists and extend it, maybe drawing a fresh link from OKF to “revenue.” In Karpathy’s version you rarely write the wiki yourself; the model does. Your job shrinks to finding and sharing new information.

The business daydream

Here Haynes shifts from describing the spec to imagining what it could become — and she’s candid that she’s speculating. Two revenue streams, in her telling. One: SEOs sell a new service, structuring a company’s knowledge into a bundle, like building a site map but far more detailed. Two, the one that “blows her mind”: experts sell their bundles directly. Buy a lawyer’s OKF, plug it into your own, and — because the bundle updates as the law changes — you’d be renting living expertise rather than a static file.

She closes on a phrase Gemini handed her: semantic unbaking. The idea that instead of grinding our knowledge into machine-readable form ourselves, we just live, learn, and share, and the format does the rest.

Key Takeaways

OKF = a folder of markdown files, one per concept, each with standard YAML front matter (type, title, description, tags, timestamp) followed by free-form body content.
The genuinely new thing is standardization, not the technology — markdown, GitHub, and YAML front matter all predate this.
A bundle includes an index file and a log file; files cross-link, forming a knowledge graph.
OKF formalizes Karpathy’s “LLM Wiki” pattern: a model that maintains a living wiki, integrating new sources rather than re-searching raw documents (the alternative being RAG).
Concepts are extracted from web pages, not one-to-one copies of them — one site might yield dozens of concept files.
A “playbook” is one file type: a process triggered by an event (e.g. a traffic-drop playbook fires when someone mentions a traffic drop).
Google’s own examples lean toward BigQuery/GA4 data, telling an agent how a business organizes its data rather than letting it guess.
Discovery is unsolved; Haynes guesses agents will find bundles via an LLMs.txt pointer.
This is a proposed v1 spec — early and likely to change.

Claude’s Take

The factual core is solid and worth knowing: OKF is real, it’s a markdown-based spec, and it formalizes a genuinely interesting idea from Karpathy about LLMs maintaining a living wiki instead of re-searching documents each time. If you strip the video down to “Google published a markdown spec for agent-readable knowledge bundles, built on the LLM-wiki pattern,” that part holds up.

The rest is an SEO consultant’s enthusiasm, and it should be read that way. Phrases like “this changes everything” and “a whole new layer for the internet” are doing a lot of work for a v1 spec with no adoption data. The buy-and-sell-your-bundle vision is pure speculation — Haynes says “I’m fairly certain” and “I don’t know this 100%” repeatedly, which is honest of her but means there’s no mechanism described, no marketplace, no pricing, no evidence anyone will do this. The “semantic unbaking” flourish is a chatbot’s turn of phrase, not an insight.

There’s also an unexamined gap between OKF-the-spec (static markdown files) and the Karpathy living-wiki vision (a model continuously rewriting those files). The video blurs them together, but a file format doesn’t make the wiki self-maintaining — that requires tooling nobody has built yet. The honest summary is: an interesting, low-stakes standard worth a weekend of tinkering, wrapped in a much larger story that hasn’t happened. Score 5 — accurate on the mechanism, heavily inflated on the implications, with the speaker’s own hedging as the tell.