Justy: Okay, Cody, this one is catnip for me. Google Cloud put out an Open Knowledge Format, and the actual pitch is basically: your agents are not dumb, they're under-briefed.
Cody: Yeah. And that's a better claim than pretending a new model fixes it. The argument is that org knowledge is scattered everywhere, so agents spend half their life reconstructing context from junk drawers.
Justy: Which is such an Exploring Next problem to be obsessed with on a Monday. Also I got here running on bad hotel coffee and one extremely fake blueberry muffin, so if I sound too excited, blame logistics.
Cody: I mean, fair. I spent this morning waiting for a package that required a signature, and of course it arrived during the six minutes I took the trash out. Anyway, this is very much the same family as all those homegrown agent docs people keep inventing.
Justy: Right. Like AGENTS dot M D, CLAUDE dot M D, random markdown vaults, little internal wikis the model is supposed to read before it does anything. Google's trying to make that shape less bespoke.
Cody: Mm-hm.
Cody: What they're actually proposing is pretty modest. Markdown files with YAML frontmatter, one required field for type, then optional structured fields so concepts like datasets, metrics, APIs, tables, and runbooks become queryable instead of just vibes in prose.
Justy: And that modesty is kind of the whole point. It's not a new knowledge platform. It's an agreement about packaging, so the same bundle can move between humans, agents, and internal tools without every team inventing their own weird folder religion.
Cody: Folder religion is exactly right. The evidence Google uses is at least concrete, which I appreciated. They published a producer-side reference agent that walks a BigQuery dataset, drafts an O K F concept doc for each table and view, then does a second L L M pass to enrich it with citations, schemas, and join paths.
Justy: Yeah.
Justy: That detail matters because it's not just, hey, imagine a standard. They're showing one plausible way the knowledge gets made from real enterprise stuff, not by asking some poor person to hand-author five thousand files on a Friday night.
Cody: And on the consumer side they shipped a static H T M L visualizer that turns an O K F bundle into an interactive graph in one self-contained file. No backend, no install, no data leaving the page. That's a nice proof that the format can be consumed by something dead simple.
Justy: Okay okay.
Justy: That part feels product-real to me. If I'm on a team building internal analysis agents or some support copilot, I do not want the adoption story to begin with, cool, now stand up three services and a vector database and a permissions layer before lunch.
Cody: Right, Justy. Though I do think the article is a little too generous about the hard part. The hard part is not serializing knowledge into markdown plus metadata. The hard part is whether the citations stay current, whether the join paths are right, and who notices when the runbook quietly stops matching reality.
Justy: Sure.
Justy: But I don't think Google is claiming the format solves freshness. I read it more like: stop making the context problem worse with incompatible one-off conventions. If the knowledge is already drifting, at least drift in a shared container.
Cody: That's fair. And technically, keeping it loose probably helps adoption. V zero point one, open from day one, no required model or framework. If they'd over-specified this, nobody would touch it outside one vendor lane.
Justy: There is a funny alternate universe where you would be mad it was too rigid, and in this universe you're mad it's not solving governance. Continuity has ruined your whole brand.
Cody: I miss being able to contradict myself freely.
Cody: The sample bundles are also a good move. G A four e-commerce, Stack Overflow, and Bitcoin public datasets. That's not proof the standard wins, but it does make the idea inspectable instead of hand-wavy.
Justy: So who actually cares… I think teams already building agents inside companies. Data teams, platform people, analytics, whoever keeps getting asked why the assistant knows the table name but not what the metric means. This is not a consumer feature. It's plumbing for org memory.
Cody: And I could see coding agents using the same pattern too, especially where repo docs, service maps, and runbooks all live in different places. Not because O K F is magic. Just because shared shape reduces glue code, and glue code is where these systems get gross fast.
Justy: Yeah. If this becomes anything, it's because enough tools agree that readable files with a little structure beat another sealed knowledge silo. Which, honestly, feels almost embarrassingly sane.
Justy: Anyway, episode four ninety-four and we're out here defending file formats in your kitchen. Very glamorous, Cody.