Antigravity CLI in Google Cloud Workbench: Notebook Automation Preview

Google Cloud added the Antigravity CLI to Agent Platform Workbench in preview on November 10, 2026. On eligible instances, you can open the terminal agent from JupyterLab or launch it with agy. The integration adds notebook skills, a Google Cloud plugin and a cell magic that generates code for review.

What changed inside Workbench?

Workbench already provides a managed notebook environment where developers can use JupyterLab and its terminal. The preview adds a terminal agent inside that environment, with access to notebook-specific workflows subject to the instance’s controls. It is separate from Google’s earlier Antigravity desktop experience and its multi-agent workflow previews.

Surface	Previous workflow	Workbench preview
Terminal	Run scripts or cloud commands directly	Start agy and review proposed tool actions
Notebook	Write cells and manage kernels manually	Notebook skills can assist with notebook and kernel tasks
Inline code	Enter a Python cell yourself	%%agy generates a separate code cell for review
Administration	Control instance and notebook access	Also review instance metadata, prediction permission and model policy

See the Workbench release notes and the Antigravity CLI guide.

Run a notebook task.

Which instance can run it?

Google documents support on Debian 12 workbench-instances-2603 or later. The CLI is not installed on the older Debian 11 image. Eligible new instances have the integration enabled by default unless configured otherwise. Check the image before diagnosing permissions or reinstalling tools.

The instance service account needs permission to invoke the relevant model endpoint. Google identifies aiplatform.endpoints.predict and describes using the Vertex AI User role where appropriate. Apply permissions to the service account that runs the instance. An end user’s ability to open JupyterLab is a separate authorization decision. Choose permissions and model access consistent with your organization’s policy.

Run a small notebook workflow

Open the eligible instance in JupyterLab, open a terminal and enter:

agy

Use synthetic data. Ask for a proposal to summarize an existing CSV without making a network request or writing changes. Read the proposed action and approve only the expected operation.

To generate a cell inside a Python notebook, start the cell with:

%%agy
Generate Python code that groups a synthetic DataFrame by status and prints counts.

Google documents that this generates a new cell below. Review it before executing it. The cell magic cannot request approval and refuses actions that require it. The terminal interface is the place to review requests for permitted actions. No Workbench instance was provisioned to run this example.

How to control access and troubleshoot

• Launcher tile missing. Confirm the image and enable-antigravity instance metadata. A metadata change takes effect when the instance is stopped and started. Plan for that restart.
• Prediction denied. Verify the instance service account, the relevant permission and any organization policy controlling model access.
• Cell magic refuses a request. Restrict the request to code generation, review the result and use the terminal workflow for actions that need approval.
• agy still exists after disabling integration. Google documents that disabling metadata removes integrated features at restart while the command stays installed. Use policy and permissions to control model calls.

Keep approvals enabled and review shell commands, file writes and network requests. Files read by a tool can carry prompt-injection instructions. The instance identity is the real security boundary. Review credentials, project access, network restrictions and permitted data before involving the agent. Check the current billing terms for the instance and any model or services you use.

FAQ

Does this deploy an ASP.NET Core application? It is a notebook and terminal workflow. A .NET team could use it to analyze sanitized API telemetry or prepare a diagnostic notebook. Deployment remains a separate workflow.

Can it execute every generated notebook cell automatically? The documented cell magic produces code for review. Running it is a separate step.

Is this generally available? The November 10 release note identifies it as preview. Verify support terms and environment constraints before depending on it.

Related reading includes the Workbench metadata guide and ASP.NET PB’s AI agent guide.