Azure Skills Plugin: A New Architecture That Injects "Azure Judgment" into AI Coding Agents¶

For / Key Points

For: Infrastructure and platform engineers who have started using AI coding agents (GitHub Copilot / Claude Code) in production. Those looking to streamline Azure deployments and resource management through agents but frustrated by generic outputs.

Key Points:

Azure Skills Plugin delivers three layers in a single install: judgment (Skills) + execution (MCP Server) + AI-specific (Foundry MCP Server)
It differs from prompt packs and documentation references by handing agents the decision-making logic itself—"which service to choose"
As one of the earliest use cases of the Agent Plugin specification introduced in VS Code 1.110, it serves as a key reference point for the spread of the Skills pattern

What Happened: Agents Can Write Code but Cannot "Decide"¶

The question this section answers: Why was Azure Skills Plugin needed?

GitHub Copilot and Claude Code deliver high accuracy in code generation. But "should this app use App Service or Container Apps?", "which validations should run before deployment?", "what SKU fits this workload?"—these cloud deployment decisions fall outside the scope of code generation.

The result is the "generic tutorial" problem. Ask an agent to "deploy a Python Flask app to Azure" and you get a list of az commands and documentation links. No infrastructure code generated from project analysis. No validation execution.

On March 9, 2026, Microsoft's Chris Harris announced Azure Skills Plugin on the All Things Azure blog, designed to structurally bridge this gap between code and production environments¹.

Three-Layer Architecture: Skills, MCP Server, and Foundry¶

The question this section answers: What is the internal structure of Azure Skills Plugin?

Azure Skills Plugin deploys three functional layers from a single installation. This is the decisive difference from conventional tooling.

Layer 1: Azure Skills (Judgment Layer)¶

Over 20 curated Skill files ship with the plugin. Each Skill describes workflow definitions, decision trees, and guardrails in Markdown format, instructing the agent on "when," "what," "in what order," and "what to avoid."

Key Skills include:

Skill Name	Role
`azure-prepare`	Project analysis → Dockerfile, IaC code, and `azure.yaml` generation
`azure-validate`	Pre-deployment preflight checks
`azure-deploy`	Deployment pipeline execution via Azure Developer CLI
`azure-cost-optimization`	Cost waste detection and reduction proposals
`azure-diagnostics`	Troubleshooting via logs, metrics, and KQL queries
`azure-observability`	Monitoring setup and optimization
`azure-compliance`	Compliance checks
`azure-rbac`	Permission management
`azure-compute`	Compute resource selection
`azure-cloud-migrate`	Migration assistance

These are not mere tool catalogs. For example, azure-prepare teaches the agent an entire workflow: "analyze the project structure, select appropriate Azure services, and generate necessary infrastructure files." A veteran Azure engineer's judgment process is packaged as a Skill file.

Layer 2: Azure MCP Server (Execution Layer)¶

Over 200 structured tools spanning 40+ Azure services. Retrieving resource lists, checking pricing, querying logs, running diagnostics, provisioning infrastructure—all invoked by agents through the standardized MCP protocol.

Skills decide "what to do." MCP Server "actually does it." This separation is the architectural key.

Layer 3: Foundry MCP Server (AI-Specific Layer)¶

Connects to Microsoft Foundry (formerly Azure AI Foundry), enabling model deployment, agent management, and model catalog operations. This specialized layer serves teams building AI applications on Azure, covering AI-specific workflows that generic cloud guidance cannot address.

The design principle in one sentence: "Skills guide, MCP executes, and the Plugin maintains coherence between both."

Before/After: Same Question, Different Output¶

The question this section answers: What changes concretely?

The comparison presented in Microsoft's official announcement illustrates this shift clearly¹.

Without the plugin: Ask to "deploy a Python Flask app to Azure" and you get a generic tutorial. Fragments of az commands and documentation links—nothing more.

With the plugin: For the same request, the agent sequentially invokes azure-prepare → azure-validate → azure-deploy Skills. Dockerfile generation based on project analysis, infrastructure file creation, deployment configuration, validation execution, and actual Azure operations via MCP tool calls—all executed as a single flow.

The agent transforms from "advisor" to "executor." The qualitative difference in output is substantial.

Relationship to VS Code Agent Plugin Specification¶

The question this section answers: What technical foundation does Azure Skills Plugin stand on?

The release timing was no coincidence. VS Code 1.110, released on March 4, 2026, introduced the Agent Plugin specification as a preview². Azure Skills Plugin is one of the first large-scale plugins built on this architecture.

Agent Plugins bundle slash commands, Skills, custom agents, Hooks, and MCP Servers into a single installable package. Distributed via Git repositories, they enable version control, PR reviews, and access control.

my-plugin/
  plugin.json           # Plugin metadata
  skills/
    my-skill/
      SKILL.md           # Skill definition
      run-script.sh      # Support scripts
  agents/
    my-agent.agent.md    # Custom agent
  hooks/
    hooks.json           # Lifecycle hooks
  .mcp.json              # MCP Server definition

This structure enables Azure Skills Plugin to operate across multiple hosts:

GitHub Copilot in VS Code: Interactive operations within the editor
Copilot CLI: Terminal-first workflows
Claude Code: Anthropic's coding agent

The same Skill and MCP definitions are shared across hosts. This portability is a defining characteristic of the Agent Plugin specification.

Installation and Verification¶

The question this section answers: How do you actually set it up?

Prerequisites: Node.js 18+, authenticated Azure CLI (az login), and Azure Developer CLI authentication (azd auth login) for deployment operations³.

For Copilot CLI¶

# Add marketplace source (first time only)
/plugin marketplace add microsoft/azure-skills

# Install the plugin
/plugin install azure@azure-skills

Note that MCP Servers need to be loaded after installation.

For VS Code¶

Search for "Azure MCP" in the VS Code Extension Marketplace and install the extension (Extension ID: ms-azuretools.vscode-azure-mcp-server). GitHub Copilot for Azure auto-installs as a companion extension, loading the Skills.

For Claude Code¶

Install using the same plugin commands as Copilot CLI.

Verification¶

After installation, test with these prompts:

"List my Azure resource groups" → MCP Server calls the resource group retrieval tool, returning actual resources
"What Azure services do I need to deploy a Python Web API?" → azure-prepare Skill activates, returning project-analysis-based recommendations

The Significance of the Skills Pattern: Packaging Tacit Knowledge¶

The question this section answers: What structural change does Azure Skills Plugin demonstrate?

The essence of Azure Skills Plugin lies not in individual Azure capabilities but in the "Skills pattern" itself.

Previously, the tacit knowledge in veteran engineers' heads—service selection criteria, pre-deployment checklists, cost optimization insights—was shared through Confluence documents and team conversations. Whether developers actually read those documents was never guaranteed, and no mechanism existed for AI agents to reference them.

The Skills pattern converts this tacit knowledge into "installable artifacts that agents automatically consume." Agents don't need to read Confluence. Decision logic is encoded as Skill files and automatically loaded at the appropriate time.

Mitch Ashley of The Futurum Group characterized this shift: "Organizational knowledge that lived in senior engineers' heads becomes installable, reviewable artifacts"⁴. Teams that can package their expertise in this model will accelerate deployment cycles, while teams relying on agents' generic training will continue receiving generic answers—a polarization trajectory worth watching.

Constraints and Caveats¶

While Skills improve agent judgment, several constraints require awareness:

Agent Plugin specification is in preview: As of VS Code 1.110, the spec may change before stable release
Skills are probabilistic controls: Skills are instruction text; there is no guarantee LLM output will follow them 100%. Guardrails are mitigations, not absolute controls
Permission management demands care: Since Azure MCP Server executes actual Azure operations, the scope of Azure CLI sessions handed to agents requires equal or greater caution than traditional CI/CD pipelines

Summary¶

Azure Skills Plugin realizes the concept of "injecting cloud judgment into AI agents" through a three-layer structure: Skills (judgment), MCP Server (execution), and Foundry MCP Server (AI-specific). As one of the earliest use cases of the VS Code Agent Plugin specification, it stands as a concrete example of the Skills pattern—converting organizational tacit knowledge into agent-consumable artifacts.

At this turning point where agents shift from "generic advisors" to "entities that execute actionable workflows," the platform engineer's role shifts too. Instead of writing documentation, package Skills and design agent behavior—a journey that starts at aka.ms/azure-skills.

Chris Harris, "Announcing the Azure Skills Plugin", All Things Azure (Microsoft DevBlogs), March 9, 2026. https://devblogs.microsoft.com/all-things-azure/announcing-the-azure-skills-plugin/ ↩↩
"February 2026 (version 1.110)", Visual Studio Code Release Notes. https://code.visualstudio.com/updates/v1_110 ↩
Chris Harris, "Azure Skills Plugin - Let's Get Started!", All Things Azure (Microsoft DevBlogs), March 17, 2026. https://devblogs.microsoft.com/all-things-azure/azure-skills-plugin-lets-get-started/ ↩
Mitch Ashley, VP and Practice Lead, The Futurum Group. DevOps.com article comment, March 2026. https://devops.com/microsoft-azure-skills-plugin-gives-ai-coding-agents-a-playbook-for-cloud-deployment/ ↩