Topic 4 of 10
Agents, notebooks, and developer environments
Three items in section 3.1 of the exam guide: deploying an agent to Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine), managing notebooks in Agent Platform Workbench (formerly Vertex AI Workbench) and BigQuery, and managing developer environments such as Cloud Workstations.
Contents
Deploying an agent to Agent Runtime
Agent Runtime is a fully managed runtime to deploy, operate, and scale agentic applications without dealing with infrastructure.
- You can customize the agent's container image with build-time installation scripts for dependencies.
- Security: VPC Service Controls compliance, configurable authentication, and IAM.
- Access to models and tools, such as function calling.
- For historical reasons, the API resource is called
ReasoningEngine. - It is observed with Cloud Trace and Cloud Logging.
Workflow and deployment options
Set up the environment: A Google Cloud project, a Cloud Storage bucket as the deployment staging area, and the Agent Platform SDK for Python.
Develop the agent: With the framework you prefer.
Deploy it: On the Agent Runtime managed runtime.
Use it: Query the agent through API requests.
Manage it: List, update, and delete deployed agents.
| Option | When |
|---|---|
| Container image | From Artifact Registry. Any language or framework that meets the runtime contract |
| Dockerfile | Agent Runtime builds and deploys the image |
| Agent Platform SDK | Managed templates for LangChain, LangGraph, AG2, and LlamaIndex (agent_engines.create(...)) |
| Agent Development Kit | Full integration: the adk CLI (Python) or adkgo (Go) |
| Custom template | Other frameworks, such as CrewAI |
The Agents CLI covers the whole lifecycle (scaffold, evaluate, deploy, publish, and observe) with templates, a playground, Terraform for infrastructure, and CI/CD pipelines with Cloud Build.
Agent identity and permissions
- Whoever deploys needs the Agent Platform User role (
roles/aiplatform.user). - By default agents run as the AI Platform Reasoning Engine Service Agent.
- You can also use an IAM agent identity, tied to each individual agent, for finer access control, or your own service account.
- Permissions the agent needs on other services (for example, a database) are granted to that identity.
Agent Platform Workbench
Jupyter-based development instances for the whole data science workflow. Each instance is a Compute Engine VM with JupyterLab and preinstalled deep learning packages (TensorFlow, PyTorch), CPU-only or GPU-enabled.
Data access
Browse Cloud Storage and query BigQuery tables from JupyterLab.
Notebook executions
The executor runs a notebook once or on a schedule on Agent Platform custom training, even with the instance stopped; results are stored in Cloud Storage.
Security
Google's network or your own VPC, a VPC Service Controls perimeter, CMEK, and Confidential Computing.
Management
Idle shutdown (on by default), additional conda environments as kernels, and GitHub sync.
Notebooks in BigQuery
BigQuery integrates Colab Enterprise notebooks for data science and ML workflows in a single interface, combining SQL, Python, text, and visualizations.
- Text (Markdown), code (Python), and SQL cells; SQL cells can reference Python variables.
- BigQuery DataFrames with no extra setup, SQL autocompletion, and visualizations with matplotlib or seaborn.
- Build, evaluate, and deploy BigQuery ML models.
- Version history and sharing with IAM.
- The Data Science Agent automates exploratory analysis, cleaning, visualization, and ML tasks.
Cloud Workstations
Managed development environments with built-in security. Instead of each developer installing tools, you define a reproducible environment and access it from a browser IDE, local editors (VS Code, JetBrains), or SSH.
| Resource | What it defines |
|---|---|
| Workstation cluster | A group of workstations in a region and the VPC they attach to. Unrelated to GKE clusters |
| Configuration | The template: VM type, persistent storage, container image with the environment, IDE. Shared with IAM |
| Workstation | Each developer's environment, created from a configuration |
- They run on ephemeral Compute Engine VMs in your project and start or stop on demand.
- When stopped, the VM and runtime data are deleted; with a persistent home directory (a persistent disk) data survives between sessions.
- Idle timeout and running timeout shut down unused workstations and force restarts to apply changes.
- Configuration changes apply to each workstation the next time it restarts.
- Each workstation has its own hostname; with a private gateway, it is reachable only from your private network.
- With the base editor (Code OSS), Gemini Code Assist helps write code.
Which environment to use
| Need | Tool |
|---|---|
| An AI agent in production | Agent Runtime |
| Data science with Jupyter, GPUs, and scheduled runs | Agent Platform Workbench |
| Analyzing BigQuery data with SQL and Python | Colab Enterprise notebooks in BigQuery |
| Standardized software development environments | Cloud Workstations |
Exam scenarios
| Situation | Answer | Why |
|---|---|---|
| Take an agent built with ADK to production without managing infrastructure | Agent Runtime | A managed runtime for agents. |
| The agent is packaged as a container image | Deploy the image from Artifact Registry | Agent Runtime supports any language that meets the contract. |
| Run a training notebook every night even with the instance stopped | Scheduled execution in Workbench | The executor runs the notebook and saves the result. |
| Combine SQL and Python on BigQuery tables with visualizations | Colab Enterprise notebook in BigQuery | SQL and Python cells in one interface. |
| Every developer must have the same environment without installing anything | Cloud Workstations with a shared configuration | The configuration is a reproducible template. |
| Workstations must shut down on their own when nobody uses them | Idle timeout in the configuration | Cuts costs and applies changes on restart. |
| Workbench notebooks left running by mistake | Idle shutdown | Enabled by default. |
Common exam traps
❓ What is Vertex AI Agent Engine called now?
Agent Runtime, within Gemini Enterprise Agent Platform. In the API the resource is still called ReasoningEngine.
❓ Which identity does an agent on Agent Runtime use by default?
The AI Platform Reasoning Engine Service Agent. You can use an agent identity or your own service account instead.
❓ Is a Cloud Workstations cluster a GKE cluster?
No. It groups workstations in a region and VPC; it is unrelated to GKE.
❓ Are a workstation's files kept when you stop it?
Only those in the persistent home directory, if configured. The VM and runtime data are deleted.
❓ When does a Cloud Workstations configuration change take effect?
When each workstation restarts.
❓ What is a Workbench instance underneath?
A Compute Engine VM with JupyterLab.
Did you get this topic?
Test what you just learned
A company wants its 80 developers to work in the same environment (IDE, libraries, SDK version) without installing anything on their laptops, with environments that shut down on their own when idle and a version change that reaches everyone. What should it use?
Ready to test "Agents, notebooks, and developer environments"?
Original questions, each with a full explanation.
Start a mock exam