Microsoft just released skills that teach AI agents how to navigate its enterprise data platform — the same way you'd teach a junior engineer the company playbook.
The Summary
- Microsoft open-sourced a GitHub repo of "skills" that teach AI assistants how to work with Microsoft Fabric — its unified data platform for analytics, warehousing, and BI
- Skills are structured instruction sets covering APIs, query patterns, deployment workflows, and operational best practices across Fabric workloads
- Install them via GitHub Copilot CLI in modular bundles: authoring (APIs, notebooks, ingestion), consumption (querying, exploration), operations (diagnostics, performance), or Power BI-specific workflows
The Signal
This is infrastructure for the agent economy. Not infrastructure like servers or bandwidth. Infrastructure like training data, but operational. Microsoft is publishing the institutional knowledge that would normally live in Slack threads and onboarding docs, reformatted so AI agents can execute it.
The repo divides skills into bundles that mirror how actual teams work. Authoring bundles cover creation and management through REST APIs, CLI automation, notebooks, SQL variants, and semantic models. Consumption bundles handle read-only workflows — querying warehouses, exploring lakehouses, searching catalogs. Operations bundles focus on diagnostics and performance. You can filter by specific workloads: SQL data warehouse, Spark, Eventhouse.
"Skills are reusable AI assistant instructions for working with Microsoft Fabric."
The Model Context Protocol (MCP) tie-in matters here. MCP is Anthropic's open standard for connecting AI models to external data sources and tools. Microsoft is building skills as MCP systems, meaning any MCP-compatible AI — Claude, GitHub Copilot CLI, VSCode extensions — can use them. This isn't vendor lock-in. It's vendor bridge-building.
What Microsoft is really doing: they're teaching agents their enterprise playbook at scale. Every large company has proprietary knowledge about how to actually use their stack. Query optimization tricks. Deployment sequences that don't break things. The difference between what the API documentation says and what actually works in production. That knowledge usually transfers human-to-human, slowly, expensively. Microsoft is transferring it human-to-agent, once, in public.
The install command structure reveals the design philosophy:
- Full bundle installs everything except Power BI authoring (too specialized)
- Focused bundles let you scope by role (author vs. consumer vs. operator)
- Workload filters let you scope by technology (Spark vs. SQL vs. event streaming)
This granularity matters because agents don't need to know everything. They need to know the right things for their context. A monitoring agent doesn't need authoring skills. A data pipeline agent doesn't need consumption patterns.
The Implication
If you're building agents for enterprise tools, watch this pattern. Skills-as-code is the training wheels for AI that works inside companies. The shift isn't teaching models about your product in context windows. It's publishing structured operational knowledge that agents can install and execute.
For enterprises: the companies that document their workflows as agent-readable skills will onboard faster, automate deeper, and lose less knowledge when people leave. Microsoft just showed the format. Your internal tools are next.