Agent Platform Model Registry: Versioning, Aliases, and Evaluation

GCP Study Hub
October 22, 2026

Agent Platform Model Registry, part of Gemini Enterprise Agent Platform (formerly Vertex AI), is a centralized repository for managing machine learning model artifacts, versions, and associated metadata across the model lifecycle. It brings AutoML models, custom-trained models, and models imported from outside Agent Platform into a single view. For the Professional Machine Learning Engineer exam, it helps to know how the registry handles versioning, mutable aliases, model evaluation, and direct endpoint deployment.

Each registered model entry displays versioning details, deployment status, and model type, such as tabular, object detection, or image classification. From this unified view, teams can check how many versions exist for a model and which version is currently active by default. Models can also be organized with labels and metadata, including team ownership, deployments, and application context.

Model versioning and mutable aliases

A single registered model can track multiple versions over time, even when those versions are built with different algorithms. A churn classifier, for example, might have separate versions trained with random forest, gradient boosting, and logistic regression. Each version records its state (such as ready), description, creation time, and metadata labels showing details like who created the model and which team owns it.

One version is designated as the default, which means it is selected whenever no specific version is requested, and that same version can be tagged for production use. Teams can also assign aliases such as experimental, staging, or production to individual versions. An alias is a named, mutable reference to a specific model version. It is named because you define the string, and mutable because it is not permanently bound to the version it originally pointed to.

Because an alias can move from one version to another, promoting a new model or rolling back becomes a pointer update rather than a code change. When a newer version passes validation, you move the production alias onto it, and rolling back is the same operation in reverse. If deployment code references a hardcoded version ID, that code must be edited every time you retrain. When deployment code points to an alias instead, the code stays the same and only the alias moves. For the Professional Machine Learning Engineer exam, this distinction explains how teams decouple deployment workflows from specific version IDs during retraining and rollback.

Registering a new model version

You can register a new version of a model from the Google Cloud console, the Python SDK, or a direct REST API call. When uploading a version through the REST request body, three fields define where the model files live, how predictions are served, and how the version is grouped:

  • Artifact URI: the Cloud Storage directory containing the model artifact and any supporting files the model requires.
  • Image URI: nested under the container spec, this specifies the Docker image used as the custom container for serving predictions from the model.
  • Parent model: setting this field to an existing model's resource name tells the registry to file the upload as the next version under that model rather than creating a separate brand new model.

Grouping uploads under a parent model makes it straightforward to track model evolution over time while supporting reproducibility and governance in machine learning workflows.

Evaluating classification models before deployment

Agent Platform Model Registry provides an evaluation interface for reviewing how a trained classification model performs before deciding which version to deploy. Adjusting the confidence threshold updates key classification metrics dynamically, including F1 score, precision, and recall.

Alongside those threshold-sensitive metrics, the evaluation view presents three plots to help pick an operating cutoff:

  • Precision-recall curve: shows the tradeoff between precision and recall across thresholds.
  • ROC curve: shows the relationship between the true positive rate and the false positive rate.
  • Precision-recall by threshold chart: plots precision and recall against confidence cutoffs so you can choose a threshold suited to the use case.

Deploying versions and comparing evaluations

Once a model version has been evaluated and tested, you can deploy it directly to an endpoint from that version's options menu inside Agent Platform Model Registry. Deploying without switching contexts keeps training, versioning, evaluation, and serving connected in one workflow, which helps teams deliver models faster and keep track of what is running in production.

After deployment, Agent Platform also supports comparing up to five evaluations side by side. Teams can compare different evaluations of the same model version, such as before and after retraining or across different datasets. The comparison view includes a confidence threshold slider showing how precision, recall, and F1 score respond to threshold changes, along with core metrics for each evaluation: area under the precision-recall curve, area under the ROC curve, log loss, F1 score, precision, and recall.

The same three classification plots appear in the comparison view, covering the precision-recall curve, the ROC curve, and precision and recall across different confidence thresholds. Comparing evaluations side by side helps teams monitor model drift, measure the impact of retraining, and decide which model version to promote to production.

Our Professional Machine Learning Engineer course covers Agent Platform Model Registry alongside endpoint deployments and model evaluation, with practice questions that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow