How to Install Microsoft Agent Framework Builder Skills for .NET and Python

The install steps are simple, but the versions differ between the two languages. The announcement's blanket "use @main" advice no longer applies to both.

What the skills are, and what they are not​

Microsoft's Learn documentation describes the skills as Microsoft-maintained guidance for compatible coding agents. The builder skill helps a coding agent create or modify your application source code. Don't confuse it with runtime Agent Skills. Those are skills loaded by an Agent Framework application at runtime.

Three practical consequences follow:

• Neither skill is a NuGet or pip package. Microsoft's docs say the Python skill isn't installed by pip install agent-framework.
• Your application still needs its own SDK packages and model credentials.
• A compatible agent discovers the skill from its metadata. The docs say not to assume a universal slash command, so you prompt the agent in plain language.

Install: the version difference​

The announcement shows @main for both skills. The current Learn page is more specific.

For .NET, the Learn guide says the skill is not in a tagged release yet. You therefore have to specify the @main ref. Preview it first with gh skill preview, using the same @main ref.

For Python, the Learn page says the skill shipped in the python-1.23.0 repository tag. The unqualified command installs the latest tagged version. The page says to use @main only if you deliberately want unreleased guidance.

Be aware that search results are inconsistent here. The repository README, as indexed about 35 days ago, still says the explicit @main ref is required until a tagged Agent Framework release includes the skill. Some localized Learn pages also say a .NET skill isn't available yet. These appear to be older snapshots, and the current English Learn page, last updated 2026-10-28, is the most specific guidance. If an install fails, preview the ref first and try @main.

Prerequisites​

• A coding agent that supports the open Agent Skills specification.
• GitHub CLI with the preview gh skill commands, or Node.js 22.20 or later if you use the cross-agent skills CLI. Microsoft says GitHub CLI skill management is preview and subject to change.
• For Python, version 3.11 or later. Microsoft notes that Agent Framework 1.24 no longer supports or tests Python 3.10, although package metadata still permits installation. Some Learn snapshots describe 1.23 as still supporting 3.10, so choose 3.11 or later for new work.
• For .NET, a currently supported .NET SDK for new applications.

Scope and target agent​

The default GitHub CLI install is project-scoped. In non-interactive use it targets GitHub Copilot. The Learn page shows --agent options for claude-code, cursor and codex, plus --scope user for a user-level install. Check each installer's documentation for the clients and scopes it supports. Pick one installer, not both.

What the skills push your agent to do​

Both skills use a progressive-loading design. The agent inspects the project, then reads only the focused references it needs. The shared principle is the smallest suitable pattern:

• One agent, when tools and sessions are enough.
• A harness, when you need an operating loop.
• A workflow, when explicit stages, routing and checkpointing are requirements.

Microsoft's example is that a support assistant doesn't need a multi-agent system just because it has several tools.

The .NET skill, per the Learn guide, tells the agent to:

1. Inspect target frameworks, SDK version, package management and installed Microsoft.Agents.AI.* packages.
2. Add only the narrow set of packages the app needs, and check each package's maturity because some are prerelease.
3. Read endpoints, models and keys from configuration, and prefer token credentials over hard-coded API keys.
4. Validate with the repository's existing tests and runnable samples.

The announcement says the .NET guidance covers AIAgent, Microsoft.Extensions.AI, NuGet, dependency injection and ASP.NET Core hosting.

The Python skill covers plain agents, providers, tools, approvals, sessions and context providers. It also covers harnesses, functional and graph workflows, evaluation, observability, middleware, and self-hosting or Foundry hosted-agent development. The repository's pull request for the Python skill lists reference files for agents, harnesses, workflows, declarative construction, integrations, evaluation and observability. It also confirms that a SKILL.md entry point handles routing.

Optional companion: Microsoft Learn MCP Server​

Microsoft recommends pairing the skill with the Learn MCP Server. The skill supplies focused implementation guidance. The MCP server lets the agent search current documentation, fetch full articles and find code samples.

• Endpoint: [url]https://learn.microsoft.com/api/mcp[/url], a remote server using Streamable HTTP.
• Documented tools: microsoft_docs_search, microsoft_docs_fetch and microsoft_code_sample_search.
• Microsoft describes the endpoint as public and unauthenticated.
• Microsoft also says it isn't an API in the traditional sense. Tools and their request and response formats may change, so clients should list tools each time they initialize the server.

The MCP server is optional. You don't need it to install either skill.

If your app also touches Azure or Foundry, the announcement points to the separate Azure Skills Plugin. That plugin has skills named microsoft-foundry, azure-prepare, azure-deploy and azure-diagnostics. It also bundles the Azure MCP Server and Foundry MCP. Azure operations still require the right credentials and permissions.

Prompting tips​

Microsoft's two sample prompts are the ones in the announcement:

• .NET: add an Agent Framework support assistant to an ASP.NET Core app. Isolate conversations per user, require approval before changing tickets, and test tool behavior with a fake IChatClient.
• Python: turn an existing research agent into a research-and-review workflow. Preserve its response format, add checkpointing, and test resume behavior with the project's existing test tooling.

The docs say to include your model provider, hosting target, persistence requirements, security boundary, approval boundaries and target runtime version when you know them. The skills tell the agent to ask about consequential unknowns, preserve existing behavior and validate changes. The Learn page also offers prompts such as reviewing an existing Agent Framework app for current APIs and package requirements.

Caveats before you trust the output​

• No production guarantee. Microsoft says the skills don't guarantee production-ready code. It tells you to review the plan, package choices, security boundaries, commands and generated changes before approval.
• Early versions. The .NET skill's metadata version is 0.1 and the Python skill's is 0.3, so expect change.
• Supply-chain hygiene. A skill is a set of instructions your agent will follow. Microsoft advises previewing content before installing and reviewing updates before trusting new scripts or instructions. That is why pinning @python-1.23.0 is attractive for teams that need reproducibility. Floating on main means your agent's guidance can change under you.
• No measured benefit yet. The announcement and documentation don't provide adoption figures or quality or productivity measurements. Whether the skills improve generated code is something to test in your own repository.
• Coverage gaps. Microsoft's Learn page says a builder skill for Go isn't available yet.

Bottom line​

For teams already building on Agent Framework, this is a low-cost way to reduce stale-API guesses and over-engineered designs from AI assistants. The practical routine:

1. Choose one installer.
2. Install the .NET skill from @main and the Python skill from the latest tag or a pinned one.
3. Preview before installing.
4. Add the Learn MCP endpoint if you want fuller documentation lookups.
5. Review everything the agent produces.

If something fails, Microsoft asks you to open an issue in the Agent Framework repository. Include the skill name, the coding agent you used, what you tried, and what went wrong.