Agents CLI on Google Cloud: Moving Agents from Prototype to Production

GCP Study Hub
November 17, 2026

Agents CLI, invoked as agents-cli, is the command-line tool for moving a Google Cloud agent project from an initial prototype to production. For the Professional Agentic Architect exam, the main distinctions come down to command boundaries. Separate command groups handle project scaffolding, infrastructure provisioning, local and remote testing, deployment, and registration with Gemini Enterprise, and each group performs only its own role.

Setting up Agents CLI and coding-agent skills

Agents CLI is designed for coding agents such as the Antigravity CLI and Claude Code. Alongside its commands, it ships with skills that a coding agent loads to decide what to do at each step of a project.

uvx google-agents-cli setup
agents-cli update

Running uvx google-agents-cli setup installs Agents CLI along with seven bundled skills covering the workflow, ADK code, scaffolding, evaluation, deployment, publishing, and observability. The coding agent loads these skills to make decisions at each step and runs the matching agents-cli commands. After setup finishes, you can verify that skills such as google-agents-cli-workflow are visible to the coding agent. When those skills change later, running agents-cli update reinstalls them.

Setup is separate from creating a project. It prepares the CLI and the coding agent once, while agents-cli create and the commands that follow work on a specific project, as the next section describes.

Moving a project from prototype to production

A project in Agents CLI generally moves through an ordered sequence of commands as it progresses from local prototyping to automated delivery:

  • agents-cli create scaffolds a new project from a template. Passing --prototype creates a minimal project, and a prototype can have its deployment target set to none. Running create again starts a new project over from a template rather than updating an existing one.
  • agents-cli scaffold enhance runs inside an existing project directory and adds deployment, CI/CD, or RAG scaffolding in place. This allows a prototype to move toward production while keeping its existing code and history.
  • agents-cli infra setup-cicd provisions the pipelines and multi-environment infrastructure once the scaffolding exists.
  • agents-cli infra single-project, run with --apply, provisions infrastructure in a single project.

Two boundaries around infrastructure and deployment are especially relevant on the Professional Agentic Architect exam. First, neither infra setup-cicd nor infra single-project registers an agent anywhere. Second, the deploy command pushes an agent to a target such as Agent Runtime and creates a deployment metadata file, without adding CI/CD scaffolding to the project. Deploying directly without pipelines leaves future releases to manual steps.

Day-to-day testing, building, and evaluation

Once a project exists, day-to-day work relies on four separate commands that handle different tasks:

  • agents-cli playground starts a local interactive web interface for trying an agent. It does not send a prompt to a remote agent.
  • agents-cli run sends a single prompt to an agent and prints the reply, without opening an interactive interface. For a deployed agent, run takes --url and requires --mode, which accepts a2a, adk, or adk_live.
  • agents-cli build builds the project.
  • agents-cli eval run runs an evaluation to measure the agent's behavior. It does not send a standalone single prompt, and it changes nothing in the project's files.

The deploy command also provides a --status flag, which reports on a pending deployment rather than talking to the agent. Keeping playground, run, eval run, and deploy --status separate avoids confusing local interactive testing, single-prompt invocations, evaluations, and deployment status checks.

Publishing a deployed agent to Gemini Enterprise

Deployment and publishing are separate steps in Agents CLI. The deploy command places an agent on a runtime, and its --deployment-target flag accepts agent_runtime, cloud_run, and gke. Making a deployed agent available to people is the job of the publish command group, whose gemini subcommand handles registration with Gemini Enterprise.

That subcommand accepts flags such as --agent-runtime-id and --gemini-app-id to identify the agent and the Gemini Enterprise app. If the agent's ID is not passed, the command reads it from the deployment metadata file that deploy creates, which is how the two steps connect. The extension commands sit outside this lifecycle entirely, since they add, list, remove, and update extensions to agents-cli itself. When reviewing Agents CLI for the exam, we would group the commands by responsibility: setup and skills, in-place scaffolding, infrastructure provisioning, day-to-day runtime and evaluation, deployment, and publishing.

Our Professional Agentic Architect practice exams test Agents CLI workflows alongside Agent Runtime deployment and Gemini Enterprise publishing, with detailed explanations that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow