Industry story
OKF v0.2 adds trust, provenance, and attestation signals for agent-written knowledge
agent-framework orchestration reliability tool-use
Agent-written knowledge at scale has a trust problem, and OKF v0.2 is Google Cloud's answer to it. When agents are generating thousands of knowledge entries overnight, downstream systems have no way to know whether to believe any of it. The new YAML frontmatter fields in v0.2 address that directly: provenance tracks what a document was built from, generated versus verified fields establish a trust tier, stale_after flags freshness, and the new Attested Computation type locks reported numbers to a sanctioned SQL query that a deterministic, non-LLM checker validates. Any rewritten query, swapped table, or added filter fails the check. The question now is whether teams actually instrument their pipelines to use these signals, or treat them as optional metadata they'll get to later.
Full analysis
Google Cloud released Open Knowledge Format (OKF) v0.2, a format for structuring the context that AI agents need (table schemas, metric definitions, runbooks) so agents can read and write it reliably. The core problem v0.2 addresses is that when agents continuously generate thousands of knowledge entries overnight (replacing human-authored docs), downstream agents and systems have no inherent way to judge whether that content is trustworthy or current. OKF v0.2 adds five families of YAML frontmatter fields to answer: what was this created from (provenance via sources), how much to trust it (generated vs. verified fields that establish a trust tier), is it still current (stale_after and status), is it the latest version (lifecycle), and was a reported number computed the sanctioned way (a new Attested Computation concept type). The attestation mechanism is notable: agents may only fill declared parameters in a computation, cannot author or edit the query itself, and a deterministic non-LLM attester mechanically checks that the SQL actually executed matches the sanctioned definition. Any rewritten query, swapped table, or added filter fails. The format remains opt-in and backward-compatible with v0.1; reference implementations, updated sample bundles, and a Knowledge Catalog (formerly Dataplex) round-trip demo are published on GitHub.
Comments