How to Use Google Agent Skills for AI Workflows on Google Cloud

Google Agent Skills are Markdown-based reusable guides that the Agents CLI consumes to generate Terraform, deployment scripts, and validation plans for building AI agents on Google Cloud.

The google/skills repository provides a structured approach to deploying AI agents through declarative skill definitions. By using Google Agent Skills for AI workflows on Google Cloud, developers can transform high-level requirements into concrete Infrastructure-as-Code without manual configuration drift. The repository ships with a Skill Engine (the Agents CLI) that parses these Markdown files and orchestrates the entire build-deploy lifecycle.

Architecture of the Google Agent Skills Framework

The framework consists of four core components that work together to translate declarative intent into executable cloud resources.

Skill Catalog – A collection of Markdown files under the skills/ directory that encode use-cases such as building agents, configuring monitoring, and deploying GKE workloads. These files serve as the authoritative, version-controlled knowledge base.

Agents CLI (ADK) – The command-line tool agents-cli that parses a skill's SKILL.md file, presents interactive checklists, and generates required artifacts. According to the google/skills source code, this component turns declarative skill definitions into actionable implementation plans.

Reference Assets – Directories such as references/ and assets/ that hold product-mapping tables and template files. These supply concrete Cloud product recommendations and Markdown templates used during generation.

Generated Artifacts – The output includes solution-architecture.md, implementation-instructions.md, and validation-plan.md, which persist recommendations from each workflow phase.

Installing the Agents CLI and Skill Catalog

To begin using Google Agent Skills, install the skill package locally and verify available capabilities.

# Install the skill package (adds all skill definitions locally)

npx skills add google/skills

# List available skills and identify the agent deployment skill

skills list | grep "build and deploy AI agents"

# → google-cloud-solution-build-deploy-agents

Executing the Four-Phase Workflow

The core skill defined in skills/cloud/google-cloud-solution-build-deploy-agents/SKILL.md implements a rigid four-phase workflow (lines 25-34): Requirements, Design, Implementation, and Validation. The CLI orchestrates each phase sequentially.

# Start the skill's interactive workflow

agents-cli run google-cloud-solution-build-deploy-agents

During execution, the CLI:

• Collects missing information from the user or uses existing context from the prompt
• Maps logical components to concrete Google Cloud services using references/product-mappings.md
• Generates architecture diagrams in Mermaid, Terraform snippets, and deployment scripts
• Validates deployments with dry-run commands and automated test suites

Because skill files are plain Markdown, they remain version-controllable and auditable without binary dependencies.

Mapping Components to Google Cloud Services

The skills/cloud/google-cloud-solution-build-deploy-agents/references/product-mappings.md file contains the mapping logic that translates logical agent components into specific Google Cloud products. During the Design phase, the Agents CLI references this file to recommend services such as Cloud Run for compute, Vertex AI for model serving, or Firestore for state management.

Generating Infrastructure as Code and Deployment Artifacts

Phase 3 (Implementation) generates concrete artifacts based on templates stored in the assets/ directory. The CLI utilizes assets/implementation-template.md to render deployment instructions.

Generating Terraform from Phase 3:

# The generated implementation-instructions.md contains ready-to-use Terraform

cat <<'EOF' > main.tf
module "agent" {
  source   = "github.com/google/agents-cli//modules/agent"
  project  = var.project_id
  region   = var.region
  model    = "gemini-1.5-flash"
  env_vars = { "API_KEY" = var.api_key }
}
EOF
terraform init
terraform apply

The solution architecture document is rendered using assets/solution-template.md, while the validation plan uses assets/validation-template.md to structure Phase 4 activities.

Validating AI Agent Deployments

Phase 4 (Validation) executes conformance checks using the plan generated from assets/validation-template.md. The Agents CLI provides specific commands for dry-run verification and quality evaluation.

# Dry-run the deployment without applying changes

agents-cli deploy --dry-run

# Test the running service endpoint

agents-cli run --url https://my-agent-run.run.app

# Evaluate agent quality against test suites

agents-cli eval run --suite basic

Summary

• Google Agent Skills are Markdown-based specifications stored in the google/skills repository that define reusable AI workflows.
• The Agents CLI (agents-cli) parses SKILL.md files and orchestrates four distinct phases: Requirements, Design, Implementation, and Validation.
• Reference assets in references/product-mappings.md map logical components to concrete Google Cloud services.
• Generated artifacts include solution-architecture.md, implementation-instructions.md, and validation-plan.md, created using templates from the assets/ directory.
• The workflow produces Terraform configurations and deployment scripts that can be validated using agents-cli deploy --dry-run and evaluation suites.

Frequently Asked Questions

What is the Agents CLI (ADK) in Google Agent Skills?

The Agents CLI (agents-cli) is the command-line interface that implements the Skill Engine. It reads Markdown skill definitions from the google/skills repository, particularly SKILL.md files, and guides users through interactive workflows. The CLI handles the translation from declarative skill descriptions into executable commands, Terraform code, and validation scripts.

How does the SKILL.md file structure work?

The SKILL.md file, such as skills/cloud/google-cloud-solution-build-deploy-agents/SKILL.md, follows a structured Markdown format that defines the four workflow phases (Requirements, Design, Implementation, Validation). Lines 25-34 explicitly enumerate these phases. The file contains prompts, checklists, and references to mapping tables that the CLI uses to generate context-specific guidance.

Can I customize the generated Terraform configurations?

Yes. The Agents CLI generates Terraform code based on assets/implementation-template.md and references/product-mappings.md, but outputs standard .tf files that you can modify before execution. After running agents-cli run, review the generated implementation-instructions.md and extract or edit the Terraform blocks before running terraform apply.

What validation methods are available for AI agent deployments?

The validation phase uses assets/validation-template.md to structure conformance checks. Available methods include agents-cli deploy --dry-run for infrastructure validation, agents-cli run --url for endpoint testing, and agents-cli eval run --suite basic for automated quality evaluation against predefined test suites.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}