Kirk builds a tool-using agent with Microsoft agent framework

Kirk introduces Microsoft agent framework through a Python demo that progresses from a model request to a tool-using retrieval agent. For context, he distinguishes two existing Microsoft projects: semantic kernel, an open-source SDK for embedding AI capabilities in applications that he describes as more production-ready and single-agent-oriented; and auto gen, a Microsoft research framework for collaborating multi-agent systems. An AI agent, in his explanation, receives a task, reasons about its next step and acts through tools.

Using UV and Cursor, Kirk installs the pre-release agent framework package and configures Azure OpenAI credentials. His first example creates an Azure OpenAI response client and generates a poem about the Toronto Blue J's. A second uses an Azure OpenAI chat client to create an agent named the Joker, which tells a Los Angeles Dodgers joke. Neither example needs a tool. He then adds a Python get weather function, with a docstring describing its purpose, and passes it to the chat agent. The agent calls it for Los Angeles weather, but Kirk stresses that the function does not contact a real weather service; its cloudy, 18-degree answer is illustrative, not a verified forecast.

Finally, Kirk replaces the mock weather tool with get info, which queries an existing Azure AI Search index containing a hockey roster. He reuses and adapts search index manager code from a Microsoft AI chat repo: the tool embeds the user's query with text embedding three small, retrieves relevant indexed chunks, and lets an Azure OpenAI chat agent use them to answer. GPT 40 is the agent's model. Asked what number Kevin Lee wears, the agent returns 13 from the Toronto rage roster; asked instead about the Toronto Blue Jays, it replies, 'Sorry, I don't know,' as its instructions require when retrieval supplies no answer. His conclusion is that attaching Python functions as tools in Microsoft agent framework is straightforward and similar to LangChain, auto-gen and semantic kernel; this is an introductory demo, not a test of retrieval accuracy or production readiness.