Problem
Many companies still believe that solving context for agents is a matter of connectors.
Connecting Drive, wiki, CRM or data warehouse helps, but it does not solve a more basic problem: how you package knowledge so that different agents understand it, exchange it, and use it without ad‑hoc translations.
Without a common format, each system speaks its own dialect. The result is more integration and less real interoperability.
Thesis
Open Knowledge Format matters because it brings the conversation back to an uncomfortable truth: a part of the agentic future will not be solved with more middleware, but with portable knowledge.
Google’s proposal is deliberately austere:
- markdown
- files
- YAML frontmatter
The operational thesis is strong precisely because of that. If context can travel as a bundle readable by humans and agents, knowledge stops being trapped in a concrete platform.
Framework
A common knowledge format for agents needs five properties:
- Readability: a human must be able to open it and understand it.
- Portability: it must travel via git, filesystem or a simple bundle.
- Minimal structure: just enough to query useful fields.
- Neutrality: not depend on a provider or runtime.
- Agentic compatibility: serve for search, retrieval and reasoning.
Mini‑case: a company wants to give several agents the same decision base about product, policies and taxonomy. If each source is exposed with a different interface, knowledge fragments. If a common bundle exists, the work shifts from translating to governing.
Measurable signal: percentage of critical knowledge that can be exported and exchanged without losing semantics or depending on a single provider.
Why it matters now
Google Cloud presented Open Knowledge Format on June 12, 2026 as an open specification that formalizes the LLM-wiki pattern in a portable and interoperable format. The initial definition is notably austere: a directory of markdown files with YAML frontmatter and small conventions for queryable fields.
The official entry itself stresses the key idea: format, not platform. Google publishes it as an open standard and claims it has already updated Knowledge Catalog to ingest OKF and serve it to its agents. Four days later, another Google Cloud publication about data agents placed OKF back into the practical stack.
That suggests something important for enterprise: knowledge for agents is beginning to be packaged as a governable artifact, not just as live access to systems.
Anti‑example
“If the agent can query the wiki, it already has context.”
Not necessarily. It may have access and still lack interoperability, portable versioning, common semantics, or the ability to move that context to another system without rewriting everything.
Protocol (3 steps)
- Choose a small domain. Policies, taxonomy, definitions or runbooks.
- Package it as a portable bundle. Human‑readable text, minimal metadata, clear version.
- Test it with more than one consumer. If it only works in one tool, it’s still not a format.
| Layer | Question | Risk if missing |
|---|---|---|
| readability | a human understands it | opaque context |
| portability | travels without friction | lock‑in |
| structure | can be queried | clumsy retrieval |
| neutrality | works outside a platform | dependency |
| consumption | multiple agents use it | false standard |
Related
- Context Supply Chain: the supply chain that decides if your AI knows how to work
- Enterprise AI Search: why internal search is becoming an operating system
- Data Contracts for AI teams: without them there is no scale
Sources consulted
Next step
Pick a knowledge base that today lives trapped in a single tool and ask a simple question: could we export it as a useful bundle for two different agents without rebuilding it entirely?
Translated from the Spanish original with AI assistance and reviewed for accuracy. Read the original in Spanish.