Every organization runs on knowledge that was never written down. Which customers count as "active" in the monthly report. The events table everyone quietly knows to ignore. The reason two dashboards disagree and no one can say which is right. That knowledge is real, it is load-bearing, and it lives almost entirely in people's heads, in chat threads, and in the memory of the one engineer who was there. Meanwhile the work — the report, the model, the agent, the decision — happens somewhere else entirely. The distance between the two is the alignment gap, and it is where both teams and AI quietly break.
The Open Knowledge Format is an attempt to close that distance. OKF is an open, vendor-neutral specification, published by Google Cloud in mid-2026, for writing down the context an organization runs on in a form any tool can produce and any tool can read. It is a modest idea aimed at an expensive problem. And the problem is older than AI.
Intent lives in one place, execution in another
Ask why a company's numbers don't match and you rarely find a broken pipeline. You find a broken agreement. A metric means one thing to finance and another to product because its definition was never written somewhere both could see. A new hire spends three weeks relearning things that were known all along but never recorded. A decision gets made on a caveat that existed only in someone's inbox.
This is the same gap that trips up an AI agent, seen from the human side. The agent can reason. What it cannot do is know your naming conventions, your deprecated systems, or which of two identical-looking metrics is the real one. Neither can your new director, your outside partner, or the team down the hall. Intent — what we actually mean, and why — sits in one place. Execution happens in another. Nothing reliable crosses the gap unless someone carries it by hand.
Bigger models won't close it
It is tempting to assume the next generation of model makes this moot. That misreads the bottleneck. A stronger model reasons better over the context you hand it. It does not develop private knowledge of your business. Knowledge that specific is not a reasoning problem waiting for scale — it is an information problem, and information has to be supplied.
The same is true of people. Hiring a sharper analyst does not recover a definition no one ever wrote down. The constraint is not intelligence at either end. It is that the context connecting intent to execution was never captured in a form anyone else could use.
The knowledge existed the whole time. It was simply never shared in a form the next person, or the next tool, could read.
The fix is unglamorous: write it down, together
Over the past year a workaround appeared independently in dozens of places, and it is strikingly low-tech. Teams started writing their institutional knowledge as plain markdown files and pointing their agents at them — a folder consulted before work begins, maintained like any other part of the codebase. You may have met it as convention files at the root of a repo, or as "metadata as code" inside data teams.
It works for a sound reason. Software is tireless at exactly the maintenance that makes people abandon wikis: keeping links current, propagating one change across many files, never losing interest. Hand that drudgery to a machine and let people curate the substance, and a knowledge base can finally stay alive instead of rotting.
The catch is that every version of this is homegrown. Two teams can both write markdown with headers and links and still produce files that cannot talk to each other, because no one agreed on what the fields mean or how the pieces connect. Portable in theory, siloed in practice. That is the whole reason a shared format matters: it is the difference between a habit and an asset.
What OKF actually is
The restraint is the point. Knowledge lives as a directory of markdown files, one file per concept, with the file's location acting as its identity. A short header holds the few fields worth querying — type, title, tags — while the body carries everything else in ordinary prose and tables. Files link to one another the way you already link anything, which quietly turns a folder into a navigable graph of how things relate. There is no runtime to stand up, no account to hold, no library you must import. It is text, in files, in version control, sitting next to the work it describes.
Because it is just versioned files, a definition can be proposed, reviewed, corrected, and adopted the way code is — and then it stands as the version everyone points to. And because the format belongs to no vendor, that shared understanding does not evaporate when a team switches tools or a key person leaves.
The alignment dividend
It is tempting to file OKF under AI tooling and move on. That undersells it. The trait that makes it useful to an agent — context written down in a neutral, shareable form — is the same trait that closes the gap between intent and execution for humans. The links hand a newcomer, or a client, a map they can follow without a personal tour. The shared definition ends the argument about whose number is right. Handing work across departments gets cheaper because the knowledge travels with it.
Agents, in this light, are simply the most demanding readers of a discipline that pays off for the people around them first. If your teams have already started keeping knowledge in markdown for their tools to read, they have done the hard part on instinct.
So the practical move is not to wait for a better model or a bigger platform. It is to decide that the context your organization runs on gets written down once, in a form everything and everyone after can read — and to treat that as infrastructure, not housekeeping. OKF is at version 0.1: a foundation, not a finished thing. But the habit it formalizes is the one that turns intent into execution reliably, at the speed the rest of the business now expects.