Integrated Gemini Enterprise Agent Platform data

After the synchronization, the Gemini Enterprise Agent Platform (formerly Vertex AI) assets are available in the specified domain.

Warning Do not move the assets to another domain. Doing so may lead to errors during future synchronizations.

By default, the assets are shown in a plain list, but you can enable a multi-path hierarchy to show it in a tree structure.

The Default Asset Status field in the capability determines the status of synchronized assets.

  • If you select No Status, newly created assets receive the first status listed in your Operating Model statuses, and existing assets keep their assigned status.
  • If you select Implemented, all assets receive the "Implemented" status.
Note For more information about AI Model Version asset statuses, go to Configure the lifecycle stages of AI Governance assets.

Synchronized metadata per Gemini Enterprise Agent Platform asset type

Note If you remove a label in Gemini Enterprise Agent Platform after a Gemini Enterprise Agent Platform synchronization, the attribute for the removed label is removed when you integrate Gemini Enterprise Agent Platform again.

This table shows the metadata for the Gemini Enterprise Agent Platform asset types. If you do not see any of the listed synchronized metadata, you can add characteristics to the layout on the asset type page.

  • The Gemini Enterprise Agent Platform integration only ingests the latest metrics from Gemini Enterprise Agent Platform models that have been evaluated and added to the model registry in Gemini Enterprise Agent Platform.
  • The integration ingests only AI agents hosted on Agent Runtime, in any framework (including ADK).
  • A Cloud Storage object used both as fine-tuning data and in a RAG corpus appears as two separate File assets.
  • A Cloud Storage-backed Discovery Engine data store's documents are ingested as File assets, grouped in one Storage Container that the data store's AI Agent Tool reads. Documents are not ingested when the data store has access control (ACL) enabled, holds more documents than the Maximum documents per data store limit, or when listing fails.
  • The latest AI Agent Version asset's Description from source system attribute shows an agent usage summary. For example, "Agent usage: 63 sessions, 12 distinct users, sessions kept 365 days." The summary is based on the 99 most recent sessions and requires no extra permissions.
Asset type Synchronized metadata Public ID
AI Base Model Description from source system DescriptionFromSourceSystem
Vertex AI Model Version Description from source system DescriptionFromSourceSystem
Framework Framework
Model Accuracy ModelAccuracy
Model Precision ModelPrecision
Mean Squared Error MeanSquaredError
Mean Absolute Error MeanAbsoluteError
Feature Importance FeatureImportance
Version Version
Supported Input Modalities SupportedInputModalities
Supported Output Modalities SupportedOutputModalities
Supported Model Customizations SupportedModelCustomizations
Any custom labels defined via the configuration. If you use custom labels, ensure that you add them to the assignment and layout on the asset type page.  
AI Model Deployment Description from source system DescriptionFromSourceSystem
Initiating User in Source InitiatingUserInSource
Creation Date in Source CreationDateInSource
Modification Date in Source ModificationDateInSource
Implemented Content Filtering ImplementedContentFiltering
Compute Configuration ComputeConfiguration
AI Agent Description from source system DescriptionFromSourceSystem
Creation Date in Source CreationDateInSource
Tool Usage ToolUsage
Modification Date in Source ModificationDateInSource
Any custom metrics defined via the configuration. If you use custom metrics, ensure that you add them to the assignment and layout on the asset type page.  
AI Agent uses / is used by AI Model Version AIAgentUsesAIModel
AI Agent has version / is version of AI Agent Version AgentHasVersion
AI Agent Version Creation Date in Source CreationDateInSource

Description from source system

 DescriptionFromSourceSystem
Version Version
AI Agent has version / is version of AI Agent Version AgentHasVersion
AI Agent Version can call / can be called by AI Agent Tool AgentVersionCallsTool
AI Agent Tool Description Description
AI Agent Tool can consult data source / can be consulted via Inference Data AIAgentToolCanReadInferenceData
AI Endpoint Access Method AccessMethod
Access Instructions AccessInstructions
Traffic Split TrafficSplit
AI Monitor Data Drift Detection Enabled DataDriftDetection
Prediction Drift Detection Enabled PredictionDriftDetection
Schedule Schedule
Alert Configuration AlertConfiguration
URL Url
File URL Url
Storage Container Description from source system DescriptionFromSourceSystem
URL Url
Storage Container contains / contained in File FileContainerContainsFile
Database    
Schema    
Table

Description from source system

 DescriptionFromSourceSystem
URL Url
Column

Description from source system

 DescriptionFromSourceSystem

Technical Data Type

 TechnicalDataType

Synchronized metadata per agent type

The Gemini Enterprise Agent Platform integration catalogs two kinds of agents: source-code agents (ADK or packageSpec) and Agent Studio agents. They expose different metadata, so what is synchronized differs by type. If an agent is missing metadata you expected, it is likely one of these type-specific cases rather than a synchronization failure.

Synchronized metadata

Source-code agent (ADK / packageSpec)

Agent Studio agent
Agent structure:
AI Agent Version Yes No
AI Agent Tool Conditional Conditional
Traceability:
Tool reads a RAG corpus or data store (inference data) Conditional Conditional
Serving and monitoring:
AI Endpoint Yes No
AI Monitor Yes Yes
Operational trust metrics Yes Yes
Basic agent information:
Description from source system Yes Yes
Creation Date in Source Yes Yes
Modification Date in Source Yes Yes
Tool usage Yes No

Conditions:

  • AI Agent Tool and inference data link: these appear only when a corpus is linked to that agent via the rag_corpus label. Gemini Enterprise Agent Platform does not expose which agent uses a corpus. The integration creates one Vertex AI Knowledge Retrieval agent per project as a placeholder. Orphaned corpora and data store tools are grouped under it, each as an AI Agent Tool called by a synthetic Knowledge Retrieval version, reading the corpus as inference data. This keeps every corpus discoverable and governed until you link it to a real agent. To link a corpus to the agent that uses it:
    • In Gemini Enterprise Agent Platform, add a rag_corpus label to the agent, value ragcorpora-<corpus-id> (the numeric ID at the end of the corpus resource name). On the next synchronization, the corpus links to that agent automatically.
    • In Collibra Platform, open the corpus's AI Agent Tool asset and add the relation to your agent version. The tool already reads the corpus, so the corpus then appears under your agent.

Operational trust metrics

The Gemini Enterprise Agent Platform integration also collects operational trust metrics for AI agents from Google Cloud Monitoring.

Asset type Operational trust metrics Token consumption metrics
AI Agent / AI Agent Version Daily pass and fail rates

Daily input and output token counts, if the agent self-instruments them
  • Like attributes, these metrics are stored internally. However, unlike attributes, these aren't accessible through the Collibra Platform APIs.
  • These metrics power the Agents dashboard, the Quality tab on AI Agent assets, and the Operational Health theme on the AI Trust Score.
  • Metrics are updated once per day.
  • Each AI Agent gets a pass/fail AI Monitor asset, named <agent> (pass/fail monitor), with no attributes. This asset is used for attaching the daily pass and fail metrics.
  • Your agent identity needs the roles/monitoring.metricWriter permission. For more information, go to Create a Google Cloud Platform connection to an Edge or Collibra Cloud site.
  • Token consumption self-instrumentation: Unlike pass/fail, there is no built-in Google Cloud Monitoring signal for token counts. Token metrics appear only if your agent code emits a custom metric:
    • Metric name: workload.googleapis.com/collibra.agent.token_count.
    • Metric kind: a monotonic counter, such as an OpenTelemetry Counter, not a gauge or up-down counter. Since the integration reads it as a cumulative value, a non-cumulative instrument causes the read to fail immediately.
    • Required labels: reasoning_engine_id, location, and direction (input or output).
    • Optional label: project_id, only if the agent writes to a different project than it runs in.
  • Malformed or incomplete emissions, such as missing labels, are logged and skipped. You can check label names/values first if no token data appears.

To explore these metrics, in Settings, under AI Command Center, switch on Operational trust for AI agents.

For more information, go to Operational trust for AI agents and AI monitoring for AI agents.