How to use the Foundry Agentic Loop agent skill

What the skill makes the agent do

An agent skill is a folder of instructions and supporting files that an AI coding agent loads when a task matches it. This skill covers building an agent solution end to end: the hosted agent, its tools and knowledge, a React and Vite front end, a FastAPI and FastMCP back end, and deployment to Azure Container Apps.

The skill has three modes. Rapid prototyping adds an agent to an existing Foundry project quickly. Spec-driven development (SDD) runs a workflow that specifies, implements, verifies and deploys a new Foundry project with azd, and you can also run one phase on its own. Improve refines the skill and its workflow using records from past runs.

The skill is not meant for isolated Azure administration or unrelated UI edits. It delegates platform operations to a separate microsoft-foundry skill.

Install the foundry-agentic-loop skill

Download the skill ZIP from the skill's page in this directory and unzip it. Then place the skill folder in one of these locations, depending on how you want to use it.

For all of your projects, put it in your user skills folder. For one project only, put it in that project's .claude/skills/ folder. In the Claude apps, you can instead upload the ZIP under Settings, then Capabilities, then Skills.

The skill can also be installed as a GitHub Copilot CLI plugin. The skill asks for your approval before it runs any plugin install, so you stay in control of what gets added.

• Personal use: unzip into ~/.claude/skills/
• Single project: unzip into the project's .claude/skills/
• Claude apps: Settings, then Capabilities, then Skills, then upload the ZIP

Trigger the skill with natural requests

You do not need a special command. Describe the agent you want and mention Foundry, Azure or an Agentic Loop request. The skill also responds to canvas actions, which are prompts that begin with [Agentic Loop canvas action].

If you already have a Foundry project, say so, because that points the skill to rapid prototyping. If you need a new project deployed with azd, ask for an SDD build. If the skill cannot tell which mode you want, it asks once and offers the two choices: rapid prototyping in an existing Foundry project, or SDD for a new project deployed with azd.

Once a mode is chosen, the skill records it in the root AGENTS.md file. Later short commands such as deploy, test and run follow that recorded mode.

• Rapid: Build a hosted agent with a React front end in my existing Foundry project.
• SDD: Build a new agent solution with a spec, implementation and an azd deployment to Azure.
• Improve: Improve the agentic loop using the runs from last week.

Know what you need before you start

The skill needs real Azure resources to finish. Without a Microsoft Foundry project and an Azure subscription, the deployment steps cannot succeed. Deployment also needs the Azure CLI (az) and azd. Companion skills and extensions are installed through GitHub Copilot CLI with the awesome-copilot plugin.

Hosted agents use Python, and the evaluation script needs Node.js. The skill runs scripts and uses the network, so review what it runs before you approve changes.

• Microsoft Foundry project and Azure subscription
• Azure CLI (az) and azd for deployment
• GitHub Copilot CLI with the awesome-copilot plugin
• Python for hosted agents and Node.js for the evaluation script

Safety notes: scripts, network and app access

The skill runs scripts and uses the network. It asks for approval before installing helper extensions or plugins and never installs extensions from a URL. It also uses managed identities and least-privilege keyless role assignments, and it keeps credentials out of browser code and Git.

Decide the app's access model from its data. If the app only shows sample, synthetic or deliberately public data, the skill defaults to no end-user sign-in. If the app shows private, sensitive or user-specific data, the skill asks for the access model before deploying and defaults to internal ingress or Entra sign-in. Anonymous access for private data requires explicit approval, and skipping that approval carries a high risk.

• Helper extensions and plugins are installed only after you approve them
• Azure settings live only in .azure/deployment-plan.md, never in AGENTS.md
• Sensitive content capture for telemetry stays off unless you approve it with redaction and retention controls

Tips for getting good results

Give the skill explicit choices. Explicit user choices override its defaults, so name your model, framework or access model when you have a preference. The skill reuses known context and asks only for decisions it cannot infer.

Use the short commands after setup. The AGENTS.md section routes commands such as run, test, add agent, add tool, add skill, deploy and status. For deploy, the skill summarizes the plan and asks before deploying the app.

Measure improvements. In Improve mode the skill runs node scripts/evals.mjs, first to index and then with --base to compare against a commit. If the evals are unavailable, the skill stops rather than guessing.