Beyond the Chatbot: Building AI Agents with Microsoft Foundry

A compelling chatbot demo can start with a prompt. A useful agent needs an application around it.

It needs a task worth completing, tools it can use, context it can retain and a way to tell whether its actions actually helped. Once it touches enterprise systems, it also needs authorization, observability and an operational lifecycle.

That is the gap my O’Reilly live course, Hands-On Microsoft Foundry, is designed to explore: moving from an interesting model response to an agent-based workflow you can build, inspect and improve.

Start with the work, not the number of agents

Imagine an internal support workflow that gathers information, checks a service and prepares a recommended response. The interesting design question is not whether you can attach three agents to it. It is which steps need reasoning, which should remain deterministic and which require permission to act.

An agent’s goal describes the work. Its tools define the actions available to it. Memory and conversation state influence what it carries forward. Orchestration determines how those pieces interact.

These distinctions help you avoid a common trap: treating an agent as a prompt with several APIs attached. A tool call is a real operation with inputs, outputs, permissions and possible failure modes. A convincing final answer does not prove that the underlying workflow was correct.

From a first agent to an inspectable workflow

We begin with the architecture and agent lifecycle in Microsoft Foundry, including the roles of goals, memory, tools and orchestration. The opening demonstration introduces a goal-based agent in Foundry Studio and shows how to inspect its behavior.

From there, we move into building and operationalizing an agent. The agenda covers memory, configuration, deployment, REST API access and the use of simulation output and logs. A memory-enabled agent demonstration gives those concepts a practical setting.

The point of retaining context is not to remember everything indefinitely. It is to carry the right information across the right boundary. As you design an application, useful questions include what belongs to the current conversation, what should persist and what must never be shared between users or tasks.

The course includes more than 90 minutes of hands-on development demonstrations, alongside presentations, discussion and live Q&A. The demonstrations are a way to connect design decisions to observable behavior, rather than leaving them as architecture diagrams.

Connect tools without treating connectivity as trust

The Model Context Protocol, or MCP, is another major part of the course. We cover the roles of clients and servers, tool integration and the secure connection of agents to APIs.

MCP provides a standardized way to expose capabilities and context. It does not, by itself, decide that an agent is authorized to use a capability for every request. Tool registration, authentication and governance remain important parts of the application design.

The agenda includes a demonstration of registering an MCP-wrapped API tool, invoking it through an agent and reviewing the resulting trace. It is a concrete opportunity to connect the request you intended with the operation the system actually performed.

Know when collaboration adds value

Multi-agent design is not automatically an upgrade over a single agent. It introduces coordination, shared state and handoff decisions.

We explore modular workflows, roles such as planner and tool agents, conversation state and deployment patterns. A demonstration of a planner coordinating with a tool agent makes those responsibilities visible.

Trace inspection and telemetry help you ask better questions: which component made a decision, what tool received the request and where did the workflow diverge from the intended task? Evaluation then gives you a basis for improving behavior before expanding usage.

Who should join

This intermediate course is for architects, developers and AI practitioners who want to move beyond chat interfaces into task-oriented applications.

You should be familiar with cloud-based AI services and LLMs, understand basic REST APIs and Azure concepts, and have development experience in Python or C#. An Azure subscription or trial is optional if you want to follow along with demonstrations.

The learning goals span the agent lifecycle: understanding the architecture, integrating tools, designing collaboration and evaluating and deploying agent applications. This is an opportunity to work through the engineering questions behind the demo, not a promise that one class turns a prototype into a production-ready system.

If your next AI milestone involves taking action rather than generating another answer, join me for Hands-On Microsoft Foundry.

View the course and registration options on O’Reilly.

The course archive preserves the course reference. The details above show the next scheduled session when one is available.