The map: exam AI-500, AI-103 and this course

Nordlicht Reisen GmbH in Hamburg is a made-up tour operator. Its travel assistant from the AI-103 course answers individual questions. Now business customers come with group trips of ten people or more, and for that a team of several agents is to build a checked draft quote that a trip planner approves (M-01, M-02). This course designs, builds, evaluates, secures and delivers this team lesson by lesson, and it follows the skills list of exam AI-500 while doing so. This first lesson is the map: it says what the exam requires, what is established and what is not, what you should bring with you and how the case is set up.

Prior knowledge. The course assumes the AI-103 course, at least modules I1 and I2, and the AI track, at least modules T6 to T10, or equivalent experience. In addition come K3 (Bicep, managed identity, CI/CD) and K4 (decision notes). How this map is set up in AI-103 is shown by "AI-103, lesson 0.1"; here is only what is new for the four areas of exam AI-500. Nothing is repeated: every lesson has a section "Prior knowledge" with a reference and starts where the exam goes on.

New in this lesson. The case M-01 to M-12 and the decisions F-01 to F-08 appear together here for the first time. They are the starting point of all lessons; a lesson that departs from them writes "changes M-xx" or "changes F-xx" and gives the reason.

The requirement

Requirements for this lesson. Numbers are assumptions of the task.

No.	Requirement	Type	Measure	Weight	Source
0.1-01	For each of the four areas of the exam it is known in which module it is taught.	F	4 of 4 areas assigned	Must	Skills list
0.1-02	It is clear what is established about the state of the exam and what is open.	R	0 claimed statements without a source	Must	Exam page
0.1-03	The learning path fits the prior knowledge.	Q	1 chosen path with a reason	Must	Learners
0.1-04	It is clear what the course does not promise, including the language of the exam.	R	0 open questions about Python, exam language and promises	Must	Scope of the course

The exam and its state

The page "Study guide for Exam AI-500: Designing and Implementing Multi-Agent AI Solutions" on Microsoft Learn describes the exam. As its audience it describes an expert-level professional who designs, builds and optimizes scalable, production-ready multi-agent systems and who is responsible for development from design to operation. The certification is called "Microsoft Certified: Multi-Agent AI Solutions Expert" according to the plan of this course; as a prerequisite, the certification "Azure AI Apps and Agents Developer Associate" applies according to the planning of this course, that is, the one for AI-103. Open: The certification page was not read again for this lesson; read there yourself what is required today. Whoever has the ai103 course gets in professionally without a detour.

On the state, only what the pages say applies, taken literally:

Established: The "Skills measured" list is published. It has four areas with weights, which are in the next table. The exam exists only in English (Learn: "Languages: English"). Most questions concern generally available features; features in preview can appear if they are widely used.
The list carries no date. Unlike AI-103, where the list names a state, AI-500 has none. Whether and when Microsoft changes it is unknown.
Open: Whether the exam is still in a trial phase or is offered regularly is not stated explicitly by either the study guide or the exam page. The exam page says only that the practice assessment is currently not available. An answer by third parties in Microsoft Q&A is not a source this course builds on. This course therefore claims neither one nor the other. Whoever wants to take the exam reads the exam page themselves on the day of registration.

From this follows the working rule of the course: The map below applies to the list as it was read on the day of writing. If the list changes, the assignment changes; that is a task for maintenance and course management, not for you.

The four areas and the modules

Area of the exam	Weight	Prior knowledge	This course
Architect multi-agent solutions	15–20 %	AI track T6 to T8; AI-103 1.4, 1.5	Module X1, lessons 1.1 to 1.5
Develop multi-agent solutions in Azure	30–35 %	AI track T2 to T8; AI-103 2.2, 2.4, 2.5	Module X2, lessons 2.1 to 2.12
Evaluate, optimize, and monitor multi-agent solutions	20–25 %	AI track T9, T10; AI-103 2.6, 2.8	Module X3, lessons 3.1 to 3.5
Secure, govern, and deploy multi-agent solutions	20–25 %	AI track T6; K3 3.5 to 3.7; AI-103 1.4	Module X4, lessons 4.1 to 4.7

Module X0 (this lesson and 0.2) is orientation, module X5 the capstone with the group-travel team as a design and sketch. The course has 33 lessons and 52 hours. The hours do not follow the weights of the exam but what is new after AI-103: Developing multi-agent systems needs the most room.

Two areas of the exam are not in the prior knowledge but are new for everyone: the interplay of several agents (patterns, protocols, memory between agents, control of launches) and the assignment of costs and traces per tenant. Both come up in every lesson.

The case: the group-travel team

The starting point from AI-103 (R-01 to R-12) continues to apply unchanged; this course says "AI-103, R-03" and does not repeat it. New is a group-travel team. Its starting point is called M-01 to M-12:

M-01, M-02: Requests for ten people or more arrive by form and email. The goal is a checked draft quote that a trip planner (a person) approves. It never goes to a customer without approval.
M-03: Seven prompt agents in the project gruppenreisen: planer, unterkunft, transport, kalkulation, recht, texter, pruefer.
M-04: Tools on three MCP servers: buchungen (read-only), lieferanten, rechner.
M-05: Autonomy levels A0 to A3. A3 (sending, payment, cancellation) always stays a person.
M-06 to M-08: Knowledge from 120 documents and 40 supplier contracts; separation by mandantId; data in the EU, sign-in without keys.
M-09 to M-12: Environments dev, test, abnahme, prod in rg-ai500-<umgebung>; control on the Container App app-gruppenreisen with the Service Bus queue anfragen; costs per tenant; a test set with 60 cases, 10 of them malicious; all data made up.

The decisions F-01 to F-08 say how the system is built: control in code with the Microsoft Agent Framework (F-01), MCP servers behind API Management (F-02), three memory tiers (F-03), identity at the tools where Learn names no identity of its own per prompt agent (F-04), guardrails at four intervention points (F-05), evaluation from test set, human review and samples (F-06), blue/green for the app and canary for agent versions (F-07). The numbers of the case are assumptions of the task, not specifications from Microsoft.

Decision: the learning path

Who you are	Path	Limit that tips it
AI-103 and AI track done	Module X0, then X1 to X5 in order	If Azure experience is missing, first AZ-104 lesson 0.1 and K3
Only AI-103, not the track	Read the named track lesson first for each lesson	If the breadth is missing, the track first
Goal: the exam	As above, plus practice Python yourself	If the list changes, read the map again
Only one area, for example security	X0, then the module of the area	X3 and X4 build on 1.5 and 2.8: those first

Decision note. Context: You have AI-103 and want to understand systems made of several agents. Decision: The path in order, with X0 first. Consequences: 52 hours; lesson 0.2 sets up the practice account for multi-agent runs, on which later exercises build. Alternatives: Reading modules individually; rejected, because in every lesson the case builds on the earlier ones.

Python and the language of the exam

On the audience, Learn writes that whoever takes the exam should be "proficient in Python" and familiar with Microsoft Agent Framework, Model Context Protocol (MCP), retrieval-augmented generation (RAG) and LangGraph. This course teaches .NET, our stack. Lesson 2.12 shows how to read and place LangChain, LangGraph and Hugging Face Transformers and what the same task looks like with the Agent Framework in C#. Whoever wants to take the exam practices Python on their own in addition; a bridge course is not planned. Reading Python in general: "AI-103, lesson 0.3".

Verification status

This lesson contains no code for Azure. What was read were the study guide for AI-500 and the planning of the course; that was the state on the day of writing, without a date on the list itself. Nothing was run against Azure, because there is nothing to run. Open: whether the exam is still in a trial phase or is offered regularly; whether and when the list is changed; which preview features questions concern.

Exercise 1

Without looking at the table, assign six professional skills to a module and a lesson: (a) define autonomy levels and personas; (b) build an approval into the flow; (c) design a memory with lifetime and tenant separation; (d) use the AI Red Teaming Agent; (e) place LangGraph; (f) charge costs back per tenant. Afterwards compare with the solution.

Solution 1

(a) Module X1, lesson 1.1 (prior knowledge: AI track T6.4). (b) Module X2, lesson 2.9 (prior knowledge: AI track T7.3). (c) Design in X1, lesson 1.3; built in X2, lesson 2.4 (prior knowledge: AI track T7.5). (d) Module X4, lesson 4.3. (e) Module X2, lesson 2.12. (f) Module X3, lesson 3.5; the preparation for it, the tenant identifier in traces and costs, is already in lesson 0.2. Rule of thumb: What you know from the AI track and from AI-103 gets the multi-agent view here; the lesson names it under "Prior knowledge".

Common reasoning errors

"Whoever has AI-103 has AI-500 under control." AI-103 teaches single agents; exam AI-500 asks about design, protocols, memory between agents, evaluation as a whole, securing and delivery.
"The course prepares for the exam, so it is enough on its own." It promises no pass; the exam expects Python experience, and whoever takes it practices that in addition.
"The list is bindingly dated." It carries no date. Read the page again on the day of registration.
"The numbers of the case are specifications from Microsoft." They are assumptions of the task; each lesson names limits and quotas with a source.

What you learn in this lesson