Google Cloud's Knowledge Catalog now learns from agent mistakes to fix business context

What Google announced

As Google Cloud reported on its own blog, it has launched two preview features in Knowledge Catalog, which it calls a context engine for agents. One is self-learning context enrichment. The other is a redesigned interface called Knowledge Studio.

"Context" here means the company-specific knowledge an AI agent needs to answer well: how a metric is defined, which tables hold which data, and how those tables connect. Google says these features will reach customers over the next few weeks, and interested companies can request earlier access.

The problem: agents misread your business

Google's example is a simple question to a Gemini agent: what was net revenue in EMEA last quarter? To answer correctly, the agent has to know how the company defines net revenue, where the booking data sits, and how the tables join. That knowledge usually lives in several places, such as a finance playbook in Google Docs, a data model in GitHub and a Confluence page.

The traditional fix is for experts to write the definitions down by hand. Google's point is that they rarely have time to document everything. When finance changes how partner credits are handled, the written definitions go stale, and any agent relying on the old formula starts giving wrong answers.

How the self-learning loop works

The tool connects to structured data, like tables and schemas, and to unstructured sources: Cloud Storage, Google Workspace, GitHub, Confluence and SharePoint. It links the documents to the tables so an agent can find the right data and understand it. On demand or on a schedule, it drafts business definitions, table and column descriptions, join relationships, golden queries (vetted example queries) and task-specific instructions for agents. Each comes with source citations and a confidence score.

The "self-learning" part is that it also reads interaction traces from Gemini agents, Conversational Analytics in BigQuery, or a company's custom agents. Personal information is redacted. It looks for signs of trouble: ambiguous column lookups, metrics with no grounding, missing definitions, or users asking the same question in different words. It then looks up the authoritative answer and stages a fix so the error does not repeat.

On each run it also checks what changed in the sources, retires obsolete definitions and flags conflicts for review.

Where humans stay in charge

Google stresses that automation is optional. Teams can turn auto-publishing on or off and choose the confidence score, for example 85%, at which changes are approved automatically. Anything below the line, plus sensitive updates such as deletions, becomes a change request that a data steward can approve, edit or reject. Teams can also supply their own test set to check proposed updates. Every change is logged with its origin and can be reverted.

Knowledge Catalog also uses context sets, which are curated slices of the larger context graph (the map of definitions and relationships) for a particular team, agent or user group. A sales agent can use a Sales set while a finance assistant uses a Finance set, so "net revenue" can mean different things in each. Access is managed through IAM, Google Cloud's permissions system. Sets build on the Open Knowledge Format, so they can be imported and exported.

Knowledge Studio and how to read the claims

Knowledge Studio is the interface for the people doing this work. It offers a visual map of a company's data across clouds, a way to ask questions in plain language and see how agents reason over the metadata, and a place to review proposed changes with version control similar to software tools. It also tracks each enrichment run's schedule, results, token usage and duration.

Google says Knowledge Catalog makes agents nearly 4X more accurate, with up to 33% lower latency and up to 75% fewer tokens, based on internal benchmarks. Tokens are the chunks of text a model reads and writes, and they drive cost. These are vendor numbers, and the post does not describe the test setup. The post also does not mention pricing. Treat the figures as a reason to run your own trial, not as a result.