
Agent Platform Vizier, part of Gemini Enterprise Agent Platform (formerly Vertex AI), is a service for finding the best configuration for a workload. It automates hyperparameter tuning and complex optimization tasks so teams do not have to guess values such as learning rates or batch sizes and rerun training repeatedly. For the Professional Machine Learning Engineer exam, the central architectural distinction is that Vizier is a suggestion service rather than an execution engine, though it also powers hyperparameter tuning jobs when you want Google Cloud to run the trials for you.
Vizier searches a defined parameter space on your behalf. You tell the service what metric you are optimizing for and what ranges each parameter can take, and Vizier suggests a set of trial values to test. It does not execute the workload or run the trials itself. Instead, you run the workload with the suggested values and report the measured results back to Vizier.
Because Vizier never executes the workload directly, its use extends beyond machine learning model training. It performs optimization tasks for both ML and non-ML workloads across multi-cloud environments. As long as you can report a metric back to the service, Vizier can optimize the configuration even when that workload runs outside Google Cloud.
Using Agent Platform Vizier directly follows an iterative feedback loop between your environment and the service:
Every completed trial gives Vizier more information to work with, so each round of suggested values is a better guess than the one before. The loop continues until it reaches the trial limit you set or until you stop it yourself.
If you would rather not run trials yourself, Vizier also powers hyperparameter tuning jobs on Agent Platform. These jobs pair Vizier with serverless training so that each trial runs as a serverless training job and parameter values adjust based on each trial's results.
Both options use the same optimization engine underneath. The difference is whether you handle trial execution and the feedback loop yourself or let Agent Platform run them for you. For Professional Machine Learning Engineer exam scenarios, we would lean toward hyperparameter tuning jobs when pairing optimization with serverless training on Agent Platform, and standalone Vizier when a workload is non-ML, runs across multi-cloud environments, or requires you to run the trials yourself.
Our Professional Machine Learning Engineer course covers Agent Platform Vizier alongside serverless training and hyperparameter tuning jobs, with practice questions that drill these distinctions.