Google Cloud previews self-learning context tools for AI agents

By Elena Vasquez

Google Cloud says its Knowledge Catalog made AI agents nearly four times more accurate in internal benchmarks, with up to 33% lower latency and up to 75% fewer tokens.

On November 7, 2026, Google Cloud introduced self-learning context enrichment and a redesigned Knowledge Studio in preview for Knowledge Catalog. The company says both features will roll out to customers over the coming weeks.

The release targets the business context used by AI agents. Google Cloud says organizations need current definitions, rules and relationships across internal data and documents. Its product description calls Knowledge Catalog “a context engine for agents.” It links structured metadata, including tables and schemas, with unstructured material such as playbooks and internal documentation.

Google Cloud says the aim is to reduce manual work for context engineers and data stewards. That pressure grows as companies move from a small number of agents to much larger deployments. For buyers, the stakes are reliable context and governed reuse, because business logic often sits outside any single data model.

What problem is Google Cloud trying to solve?

Google Cloud says business logic is often spread across documents, source code, wikis and data models. As definitions change and data estates grow, companies struggle to maintain one trusted source for automated systems.

The enrichment feature starts with the way users and agents interact with data. Google Cloud says it analyzes those interactions to identify missing or outdated context. The system then looks for evidence that an agent needed help or produced an unreliable answer.

According to Google Cloud, the system looks for ambiguous column lookups, missing definitions, ungrounded metrics and repeated user rephrasings. Each signals that context may be incomplete, unclear or out of date.

It stages proposed fixes for review. Approved changes can then be reused across other agents in the organization. The goal is to turn corrections made during daily use into reusable context for future work.

Google Cloud states the objective is “to shift that work from manual upkeep to a feedback loop based on how agents and users interact with data.” The proposal remains subject to the governance controls built into Knowledge Studio.

Where does the enrichment system draw context from?

Google Cloud says the self-learning system can connect to structured data from first-party and third-party sources. It can also read unstructured content stored in Cloud Storage, Google Workspace, GitHub, Confluence and SharePoint.

That reach lets the system draw proposed updates from business documents, code repositories and other sources. Google Cloud positions this combination as a way to maintain context across both formal data structures and the documents that explain them.

The announcement lists several types of proposal:

• Business definitions for shared terms.
• Table and column descriptions.
• Join relationships between data sets.
• Query examples.
• Task-specific instructions for agents.

Every proposed update comes with source citations and a confidence score. The citations show where the proposed definition or relationship came from. The score gives reviewers another signal for deciding whether an update needs closer scrutiny.

The enrichment process can run on demand or on a schedule. Google Cloud says it can detect changes in source material, retire obsolete definitions and flag conflicts for review.

That process matters because context can expire silently. A column can change meaning, a policy document can replace an older procedure, or two sources can begin describing the same metric differently.

How does Google Cloud find the gaps?

Google Cloud says the system can analyze interaction traces from Gemini agents, Conversational Analytics in BigQuery and custom agents. Those traces provide evidence about where agents struggled and which information users had to supply repeatedly.

The company says the process redacts personally identifiable information. This lets the feedback loop draw on agent activity without treating personal data as enrichment material, according to Google Cloud.

The redesigned Knowledge Studio provides a central workspace for people responsible for curating and governing context. Its new interface combines a visual context graph with conversational exploration of data relationships.

Google Cloud also added a review process for proposed changes before publication. That process places a human checkpoint between automatic discovery and organization-wide reuse, while still allowing customers to configure some automation.

How does Knowledge Studio keep changes under control?

Customers can choose whether proposed updates publish automatically. They can also set the confidence threshold required for auto-approval, giving them control over how much certainty the system needs before acting alone.

Changes below that threshold are routed as change requests. Sensitive updates, including deletions, also go to a steward for review. The steward can edit, approve or reject each request before it reaches the wider context graph.

Every change is logged and can be reversed, according to Google Cloud. The company also says customer data is not used to train its models.

These controls address the main risk in an automatic enrichment system. A useful suggestion can still encode a bad definition, an incorrect relationship or a rule that only applies to one team.

How do context sets separate departments?

Context sets are curated subsets of the wider context graph. Google Cloud says a set can serve a specific team, domain, user group or agent while remaining connected to the organization’s broader catalogue.

The design lets companies maintain different definitions for similar terms. Sales and finance, for example, can retain separate meanings for a shared business term while both remain manageable within one Knowledge Catalog.

Identity and access management rules control each set. Permissions determine which users and agents can read, use or edit it. Agents retrieve context only from sets they are authorized to use.

Access to the underlying data still depends on existing permissions. Google Cloud says the context layer does not grant an agent broader access to the systems holding that data.

What can customers monitor?

The redesigned Knowledge Studio includes monitoring for enrichment runs. Users can track schedules, results, token usage and run duration, according to Google Cloud.

Google Cloud also added email notifications when a run completes. Those alerts give context teams a direct signal that new proposals, conflicts or retired definitions may be waiting for review.

The operational view is part of the larger scaling problem. Google Cloud says businesses need to keep organizational knowledge current even as definitions change and data estates expand.

What do the internal benchmarks show?

Google Cloud’s internal benchmarks found that Knowledge Catalog made AI agents nearly four times more accurate. The company also recorded up to 33% lower latency and up to 75% fewer token use compared with its test baseline.

The announcement presents those figures as evidence that better-maintained context can improve agent results while reducing work and resource consumption. The figures remain Google Cloud’s own internal performance results rather than an independently reported evaluation.

The metrics connect three parts of the product’s pitch. Better context aims to improve accuracy. Reusing approved context across agents can reduce repeated work. Lower latency and token consumption address the operating demands that come with larger agent deployments.

When will the features reach customers?

Both self-learning context enrichment and the redesigned Knowledge Studio remain in preview. Google Cloud says it will roll both features out to customers over the coming weeks.

That timing makes the next stage a test of trust. As proposed definitions reach production systems, their value will depend on whether context teams trust the suggestions, set effective thresholds and can reuse them without spreading outdated rules across agents.