An LLM can generate responses from the context it receives, but on its own it cannot query your application data, call a business API, update a ticket, or complete a task through a sequence of real-world actions. An AI agent is the system built around the model: instructions and context, state, available tools, and an execution loop that lets the model choose actions and observe their results.

In this practical, demo-driven session, we break down what changes when an LLM is placed inside an agent runtime and follow a tool call from beginning to end. We examine how tools are presented to the model through names, descriptions, and input schemas; how the model selects a tool and produces structured arguments; how the application validates, authorizes, and executes that request; and how the result is returned to the model for its next decision.

We also explore where tool calling can go wrong: selecting the wrong tool, producing invalid arguments, repeating an action, or triggering unsafe side effects. We discuss the controls that belong in the application layer, including permissions, validation, retries, timeouts, idempotency, human approval, and observability.

A live C# demo using Microsoft Agent Framework makes the complete agent loop visible in code and shows that the model requests actions—but the application remains responsible for executing and controlling them.

You will leave with a clear mental model of the difference between an LLM and an AI agent, how tool calling actually works, and how to design tools that agents can use safely and reliably.

Format: 45–60 minute technical lecture with a live C# demo and code walkthrough.

Target audience: Software developers, AI engineers, and software architects who use LLM APIs and want to understand how agent runtimes turn model output into real actions.

Level: Intermediate (Level 200). Basic familiarity with APIs and LLM concepts is helpful; prior experience with an agent framework is not required.

Demo implementation: Microsoft Agent Framework with C#/.NET. The underlying concepts are framework-independent and also apply to LangChain, LangGraph, and other tool-calling runtimes.

Earlier version: An earlier version of this session was presented as “Building AI Agents with Semantic Kernel” at Global AI Bootcamp 2025, Microsoft Reactor Tel Aviv, on March 17, 2025.

Technical requirements: Stable internet connection for the live model and tool-calling demo, plus a projector or large screen suitable for displaying code clearly.