Google proposes Open Knowledge Format for agent context
Google Cloud introduced OKF, a markdown-and-YAML format for portable agent context. It targets data teams that need AI agents to share metadata without a proprietary catalog lock-in.
The next fight in enterprise AI may be less about the model and more about the files it reads before answering. Google Cloud introduced the Open Knowledge Format, or OKF, as a draft, vendor-neutral way to package business context for agents using ordinary markdown files and YAML frontmatter.
That sounds modest, but it targets a real bottleneck. As companies move from chatbots to agents that inspect datasets, explain metrics, operate runbooks or answer domain-specific questions, the relevant knowledge is usually scattered across catalogs, wikis, notebooks, code comments and the heads of senior engineers.
OKF tries to make that context portable: a directory of markdown concept files, optional index.md and log.md files, standard links between concepts, and a small frontmatter surface where type is required and fields such as title, description, resource, tags and timestamp can be queried.
Google's strategic angle is interoperability rather than another hosted database. OpenAI, Anthropic and Microsoft are all pushing agents toward longer-running work with tool use and memory; Google is leaning on a format that can sit beside code in git, be reviewed in pull requests, and be consumed by Gemini, ADK agents, LangChain-style stacks or a plain search index without forcing teams into one vendor's catalog. That is a practical Google Cloud move: make the knowledge layer easy to adopt, then let BigQuery, Dataplex, Knowledge Catalog and agent tooling meet it where enterprise data already lives.
The v0.1 spec is intentionally small. A concept can describe a table, metric, API, playbook or business process. Its file path acts as the concept ID, the body carries human-readable context, and normal markdown links turn a folder tree into a knowledge graph that agents can traverse progressively instead of loading every document at once. Google's reference implementation includes a BigQuery enrichment agent, a static HTML visualizer, and sample OKF bundles for GA4 e-commerce, Stack Overflow and Bitcoin public datasets in the Knowledge Catalog GitHub repo.
A minimal OKF concept looks more like documentation than infrastructure:
---
type: BigQuery Table
title: Orders
description: One row per completed customer order.
resource: https://console.cloud.google.com/bigquery?p=acme&d=sales&t=orders
tags: [sales, revenue]
timestamp: 2026-05-28T14:30:00Z
---
# Schema
| Column | Type | Description |
|---|---|---|
| `order_id` | STRING | Globally unique order identifier. |
| `customer_id` | STRING | FK to [customers](/tables/customers.md). |
For builders, the immediate path is to read the OKF v0.1 specification, try the reference bundles, or adapt the Knowledge Catalog demo code. Google says Cloud Knowledge Catalog can now ingest OKF and serve it to agents, but OKF itself does not require a Google Cloud account, SDK, runtime or schema registry.
The caveat is that OKF is still a draft standard, not a mature ecosystem. Google has published the format, reference tools and examples, but adoption by third-party catalogs, governance workflows, security scanners and agent frameworks will determine whether it becomes a lingua franca or another useful convention. Pricing is mostly the wrong question for the format itself; the costs sit in whichever catalog, BigQuery project, agent runtime or review workflow a team uses around it.
For GotGemini readers, the useful takeaway is simple: if your Gemini or agentic workflow keeps rebuilding the same RAG context, OKF is a clean candidate for treating that context as versioned source material rather than prompt glue. That makes the launch small in surface area but meaningful in direction: Google is trying to standardize the handoff between enterprise knowledge and AI agents before every platform hardens its own silo.
References:
Sections marked "not disclosed" reflect omissions in the source. We don't pad them out.