From Bare AKS to Running LLMs: AI Runway Setup Skill

Where this fits

You have an AKS cluster sitting idle and a team that wants to serve an LLM on GPUs. The gap between a bare cluster and a running vLLM or KAITO deployment is packed with steps: verify node pools, install the AI Runway controller, check GPU availability, configure providers, and finally deploy a model. This skill gives an agent a concrete playbook for that journey, so a developer can say "set up AI Runway on this cluster" and get a working inference endpoint instead of a pile of TODO notes.

Why agents benefit

• Cost-aware sequencing: The skill flags expensive GPU node pools before provisioning, so agents can confirm tradeoffs instead of silently burning A100 hours.
• Verification before action: Agents check cluster state and existing resources first, reducing the risk of overwriting live configuration.
• Handles multi-step complexity: Controller install, GPU assessment, provider setup, and first deployment are broken into discrete steps an agent can execute reliably.
• Resumable execution: The skip-to-step N argument lets agents pick up an interrupted workflow without starting over or asking the user to repeat context.
• Operational handoff: If no cluster exists, the skill tells the agent to use the azure-kubernetes skill first, keeping workflows composable rather than monolithic.

Practical scenarios

Onboarding an existing AKS cluster for GPU inference

A platform engineer wants to reuse an existing AKS cluster for AI workloads. The agent verifies the cluster, assesses whether GPU nodes are present, and either enables the GPU node pool or clearly reports what needs to change before model serving is possible.

Deploying a first LLM with KAITO or vLLM

A developer wants to serve an open-source model but doesn't want to hand-write YAML for controllers, providers, and inference runtimes. The agent installs the AI Runway controller, sets up the provider, and triggers the first deployment, then confirms the model is reachable.

Resuming a setup after a failed step

A previous run died during provider configuration. The user invokes the skill with skip-to-step to continue from the failure point, and the agent revalidates the cluster state before moving forward instead of blindly rerunning everything.

Add it to your agent workflow

Install the skill from the Microsoft GitHub Copilot for Azure repository:

npx skills add microsoft/GitHub-Copilot-for-Azure --skill airunway-aks-setup

Then trigger it naturally in your agent prompt:

Set up AI Runway on my AKS cluster and deploy a small LLM for testing.

The agent will follow the skill's sequence, check costs, and report progress. You can see more context about this skill on OpenAgentSkill before deciding to adopt it.

Compare before adopting

Ask how this skill compares to alternatives in three areas:

• Quality signals: Does the repo have recent commits, open issues, and examples that show real use? This skill comes from Microsoft's official Copilot for Azure repo, which is a strong signal.
• Maintenance freshness: Check whether the README and step definitions match current AKS and KAITO versions. Fast-moving infrastructure means stale skills fail quietly.
• Workflow fit: Does your agent already have cluster provisioning or Terraform skills? This skill hand offs to azure-kubernetes, so make sure that skill is installed too for a complete path.

Why it is worth tracking

At 248 stars and maintained under Microsoft's GitHub-Copilot-for-Azure repository, this skill has credible backing and clear community visibility. The real value is its attention to cost and verification, not just command execution. Track it if you routinely deploy AI workloads on AKS or want your agents to handle GPU infrastructure responsibly. Re-evaluate it whenever AKS or KAITO releases major updates, and test it against a staging cluster before trusting it in production.