Course Overview
FROM AGENTIC IDEAS TO PRODUCTION JUDGMENT
Multi-agent systems involve much more than prompting a model. Agents use tools, retain state, exchange information, cross trust boundaries, and make decisions that can trigger real actions.
That creates architectural questions that we have to answer clearly. Which tasks need an agent? When is a workflow safer? Who owns shared state? How should permissions follow identity? What must be validated before a tool runs? How do you evaluate a variable system, recover from failure, and deploy changes safely?
Microsoft AI-500 focuses on the judgment required to design and implement production-grade multi-agent solutions. This course develops that judgment while covering the certification domains, exam framework, prerequisite, testing process, and study priorities.
You will move from LLM and agent foundations into architecture, Azure development, evaluation, monitoring, security, governance, deployment, and final review. The goal is not simply to memorize services or patterns. It is to understand why one design fits a requirement better than another.
MODULE WALKTHROUGH
Module 1: Foundations
You'll begin with the operating model behind LLM agents. You'll examine orchestration loops, tool requests, stopping rules, context, operational memory, RAG, MCP, and A2A;
The module then develops the production foundations of multi-agent coordination, including shared state, ownership, timeouts, idempotency, structured boundaries, least privilege, tracing, caching, Foundry projects, managed identity, RBAC, and Agent Service;
Module 2: The AI-500 Certification
Next, you'll unpack the four AI-500 domains and the capabilities each expects you to apply. You'll also review the certification pathway, AI-103 prerequisite, exam format, scoring, testing process, and official portal;
The module concludes with recurring judgment patterns, including using the lowest sufficient level of agentic capability, putting controls before context, naming boundaries, versioning the complete system, and containing failure;
Module 3: Domain 1: Architect Multi-Agent Solutions
This module develops your ability to choose among deterministic functions, bounded model calls, agents, and orchestrated workflows;
You'll examine Responses, Conversations, Sessions, persistent memory, semantic memory, Entra Agent ID, authentication flows, RBAC, and agent identity blueprints. You'll then compare Sequential, Concurrent, Handoff, Group Chat, and Magentic orchestration patterns;
The module also distinguishes inference compute from hosted-agent execution, explains agent and tool trust boundaries, and clarifies Foundry's control-plane and data-plane responsibilities;
Module 4: Domain 2: Develop Multi-Agent Solutions in Azure
Here, you'll move into implementation choices for secure and reliable Azure multi-agent solutions;
You'll compare prompting, fine-tuning, RAG, reinforcement fine-tuning, and distillation. You'll diagnose context failures such as sliding-window amnesia, summary drift, weak retrieval, and identity confusion;
You'll then work through Azure AI Search, Foundry IQ, MCP servers, Toolboxes, A2A, Cosmos DB memory patterns, Agent Framework, LangGraph, streaming contracts, and the Foundry SDK;
Module 5: Domain 3: Evaluate, Optimize, and Monitor Multi-Agent Solutions
This module shows how to collect evidence about agent quality, safety, reliability, and workflow behavior;
You'll examine the Foundry evaluator catalog, golden tests, LLM-as-judge, custom evaluators, human review, and red teaming. You also learn how synthetic and trace-based datasets support representative evaluation.
The module continues with the AI Red Teaming Agent, Agent Optimizer, production telemetry, workflow history, checkpoints, replay, failure classification, and recovery controls;
Module 6: Domain 4: Secure, Govern, and Deploy Multi-Agent Solutions
You'll then focus on the controls needed around production agentic systems;
The module covers least privilege, RBAC, on-behalf-of authority, managed identities, Key Vault, Foundry Guardrails, and custom middleware. It connects these controls to DTAP, Bicep, blue/green deployment, canary releases, automated rollback, durable execution, quota management, and chargeback;
You'll also examine multitenant RAG security, deterministic tenant isolation, private container registry access for hosted agents, and the ownership tradeoffs between local and service-managed agents;
Module 7: Reviewing the AI-500 Certification
The final module connects the domains through recurring patterns, traps, and architectural judgments;
You'll review why protocol success is not output validity, why durable identifiers should be saved before risky actions, and why secondary enrichment can run asynchronously;
You'll also examine stale authorization, stacked retries, incomplete evaluation data, unbounded fan-out, system-wide controls, lifecycle management, explicit failure paths, and limited blast radius;
LET ME TELL YOU... EVERYTHING
Some people love to know what they're getting in a package.
And by this, we mean, EVERYTHING that is in the package.
So, here is a list of everything that this course covers:
You will learn how agents choose actions while the orchestration layer executes tools, appends results, and repeats the loop;
You will learn to define stopping rules for verified success, retryable failure, persistent errors, missing information, and human authority;
You will learn how MCP standardizes tool discovery and invocation while permissions and validation continue to govern execution;
You will learn when Sequential, Concurrent, Handoff, Group Chat, and Magentic orchestration patterns fit different tasks;
You will learn how explicit ownership, status states, timeouts, idempotency, and version checks reduce coordination failures;
You will learn why shared state requires read and write rules, controlled context, conflict handling, and state consistency;
You will learn when MCP fits tool access and bounded delegation and when A2A fits ongoing collaboration between autonomous agents;
You will learn to structure action intentions, tool inputs, results, errors, and final outputs with explicit schemas;
You will learn how durable context protects verified facts, decisions, plans, and working state from context loss;
You will learn to capture trace IDs, decision points, context, and execution artifacts for workflow reconstruction and replay;
You will learn how Foundry projects connect models, agents, resources, managed identity, RBAC, and Agent Service;
You will learn the four AI-500 domains, their relative weightings, and the production capabilities they assess;
You'll learn about the AI-500 certification pathway, AI-103 prerequisite, exam format, scoring, testing process, and portal;
You will learn to choose among deterministic functions, bounded model calls, agents, and workflows based on task requirements;
You will learn why pause, checkpoint, preserved state, partial recovery, and protected side effects call for workflow orchestration;
You will learn to distinguish an Agent, Response, Conversation, context window, Memory, and Hosted Agent Session;
You will learn how persistent memory and semantic memory support different forms of continuity and retrieval;
You will learn to validate provenance, format, scope, and conflicts before promoting information into persistent memory;
You will learn how Entra Agent ID, identity blueprints, managed identities, RBAC, and Conditional Access constrain agent access;
You will learn when to use an on-behalf-of flow and when an agent should act independently through client credentials;
You will learn to separate model inference compute from hosted-agent execution for sessions, tools, files, and orchestration;
You'll learn about Standard, Provisioned Throughput, and Managed Compute options for model inference workloads;
You will learn to identify trust boundaries involving identities, credentials, instructions, context, arguments, and results;
You will learn to tier tool operations by risk and apply stronger authorization, logging, and approval as consequences increase;
You will learn how Foundry's data plane performs workloads while its control plane configures, observes, and governs execution;
You will learn to treat prompts as versioned systems with roles, priorities, permissions, context boundaries, and output contracts;
You will learn how fine-tuning, RAG, reinforcement fine-tuning, and distillation address different production requirements;
You will learn to diagnose sliding-window amnesia, summary drift, weak retrieval, and identity continuity failures;
You will learn how document RAG differs from agent RAG for external knowledge, execution history, and semantic memory;
You will learn to configure chunking, embeddings, filtering, hybrid retrieval, weighting, and reranking in Azure AI Search;
You will learn when Foundry IQ, Azure AI Search, or a custom retrieval solution provides the appropriate level of control;
You will learn to enforce caller identity, source permissions, sensitivity labels, and tenant boundaries before retrieval;
You will learn how MCP Toolboxes centralize authentication, permissions, discovery, access, and versioning across tools;
You'll learn about A2A AgentCards, Messages, Tasks, Artifacts, Extensions, durable work, and cross-boundary validation;
You will learn how Cosmos DB stores granular session state, turns, tool records, summaries, user facts, and vector embeddings;
You will learn to derive memories asynchronously with the Cosmos DB change feed and a Durable Functions pipeline;
You will learn when LangChain, LangGraph, Microsoft Agent Framework, and Hugging Face Transformers fit different needs;
You will learn to manage real time updates and final typed results as separate streaming contracts;
You will learn to use AIProjectClient, DefaultAzureCredential, named agent versions, Responses, and Conversations in Foundry SDK;
You will learn to combine general-purpose, RAG, safety, agent, rubric, and textual similarity evaluators;
You will learn how golden tests, LLM-as-judge, custom evaluators, human review, and release gates provide different evidence;
You will learn to create and version synthetic datasets from agent definitions, steering prompts, and reference files;
You will learn how the AI Red Teaming Agent tests content-safety risks, prompt injection, and XPIA vulnerabilities;
You will learn how Agent Optimizer compares prompts, skills, tool descriptions, parameters, models, and candidate configurations;
You will learn to distinguish observability, telemetry, and workflow history when diagnosing and recovering agent workflows;
You will learn how Foundry Guardrails and middleware enforce policies at input, model, tool, state, memory, and output boundaries;
You will learn how DTAP, Bicep, what-if reviews, blue/green releases, canary releases, evaluation gates, and rollback work together;
You will learn to design restore-safe workflows with checkpoints, durable IDs, idempotency, replay, compensation, and escalation;
You will learn how quota dimensions, AI Gateway project limits, response codes, project tagging, and FinOps affect governance;
You will learn to prevent cross-tenant RAG leakage through trusted identity, authorization, partitioning, and pre-retrieval controls;
You will learn the ownership and governance tradeoffs between local agents and service-managed agents;
You will learn to avoid stale authorization, stacked retries, incomplete evaluation data, and unbounded agent or retrieval fan-out;
OUR INVITATION TO YOU
Remember that you always have a 30-day money-back guarantee, so there is no risk for you.
Also, we suggest you make use of the free preview videos to make sure the course really is a fit. I don't want you to waste your money.
If you think this course is a fit and can take your knowledge to the next level... it would be a pleasure to have you as a student.
See you on the inside!
FAQ
FAQ area empty
Course curriculum
Chương 1 Course Introduction 1 bài · 3 phút
Chương 2 Foundations 7 bài · 1 giờ 32 phút
- Module Intro Học thử 2:22
- Foundations of LLM Agentic Systems Học thử 25:19
-
Foundations of Multi-Agent LLM Systems 20:11
- Foundations of Production LLM Agents Học thử 25:51
- Foundations of Foundry and AI-103 Học thử 16:24
- Module Outro Học thử 2:15
-
Key Distinctions - Foundations
Chương 3 The AI-500 Certification 6 bài · 1 giờ 1 phút
-
Module Intro 2:29
- The Four Domains Học thử 13:59
-
Exam and Testing Process (V2) 9:45
-
Website and Portal (V2) 2:32
-
The 10 Takeaways 30:35
-
Module Outro 2:09
Chương 4 Domain 1: Architect Multi-Agent Solutions 27 bài · 2 giờ 33 phút
-
Module Intro 3:45
-
Functions vs. Agents vs. Workflows 15:19
-
The Simplified Selection Heuristic
-
Sessions, Conversations, Memory and Responses 21:36
-
In Practice: Foundry SDK 3:14
-
In Practice: Dynamic Context with the Foundry Responses SDK 6:00
-
Architecting Memory 17:18
-
Wait, How do We Assemble Context Using the Agent Framework?
-
In Practice: Extending Context with the MAF ContextProvider 5:02
-
Agent Identity, Authentication and Authorization 16:50
-
I Compromise, You Compromise
-
Multi-Agent Architecture and Orchestration 15:21
-
You Were Already Doing HAX (... You Just Didn't Call It That)
-
Multi-Agent Compute 11:01
-
The Autoscaling Guide to Non-Hosted Agents
-
Agent/Tool Trust Boundaries and Risks 11:44
-
Developer Tooling You Shouldn't Skip: VS Code, The CLI, Dev Containers
-
In Practice: Foundry Toolkit VS Code Extension 2:23
-
In Practice: The Azure CLI 2:02
-
The Foundry Control Plane and Data Plane 10:02
-
Module Outro 3:41
-
Domain 1 Quiz (10 Questions)
-
Key Distinctions - Domain 1
-
Role Play: Interviewing for A Multi-Agent Architect Role
-
Intro to Udemy Labs 3:00
-
Intro to Udemy Graded Labs 4:57
-
Graded Lab: Set the Architecture Dials
Chương 5 Domain 2: Develop Multi-Agent Solutions in Azure 25 bài · 3 giờ 20 phút
-
Module Intro 3:53
-
Prompting for Multi-Agent Systems 17:55
-
Fine-Tuning Strategy for Agents and Models 20:28
-
Context Management and Failure Modes 19:53
-
Multi-Agent RAG Patterns 24:14
-
MCP Servers and Tool Ecosystems 15:08
-
MCP Servers Can Run on Almost Any Azure Compute Type!
-
In Practice: Toolboxes 4:05
-
A2A in Multi-Agent Workflows 19:22
-
Extending A2A
-
Memory Orchestration with Cosmos DB 16:24
-
Cosmos DB vs. Managed Redis: A Memorable Battle
-
In Practice: Cosmos DB and the... Semantic Reranker (!?) 3:13
-
The Right Caching, In the Right Order
-
Workflow Building: Agent Framework and LangGraph 11:53
-
Take The Orchestration, Leave the Runtime
-
In Practice: Microsoft Agent Framework vs. LangGraph 7:54
-
In Practice: MAF Orchestration Example: Handoff 5:17
-
Streaming in Agentic Systems 13:44
-
The Foundry SDK for Agentic Workflows 12:25
-
Module Outro 3:58
-
Domain 2 Quiz (10 Questions)
-
Key Distinctions - Domain 2
-
Role Play: Multi-Agent Development Mentoring Meeting
-
Graded Lab: Fix the Development Settings
Chương 6 Domain 3: Evaluate, Optimize, and Monitor Multi-Agent Solutions 24 bài · 2 giờ 48 phút
-
Module Intro 4:18
-
Evaluation: The Foundry Evaluator Catalog 17:56
-
In Practice: The Foundry Evaluator Catalog 6:16
-
An Evaluator Appears! (Two, Actually)
-
Evaluation: LLM-as-Judge And Golden Tests 17:11
-
Evaluator Consistency Means... Just Consistency
-
Evaluation: Custom Evaluators 14:51
-
In Practice: Rubric and Custom Evaluators 5:45
-
Evaluation: The Three Evaluation Modalities 15:43
-
A Known Bias
-
Synthetic Data Generation 17:08
-
In Practice: Synthetic Data Generation 3:13
-
The AI Red Teaming Agent 16:25
-
In Practice: The AI Red Teaming Agent 4:08
-
The "Usual" Test Types, Revisited
-
The Agent Optimizer 15:01
-
In Practice: The Agent Optimizer 3:34
-
Production Telemetry and Failure Modes 22:31
-
For the Agent Framework Coders: The Importance of Connection Strings
-
Module Outro 3:32
-
Domain 3 Quiz (10 Questions)
-
Key Distinctions - Domain 3
-
Role Play: Mentoring Meeting on Evaluation
-
Graded Lab: Set the Evaluation Dials
Chương 7 Domain 4: Secure, Govern and Deploy Multi-Agent Solutions 24 bài · 2 giờ 37 phút
-
Module Intro 3:09
-
Identity, RBAC and Access Control 14:47
-
On-Behalf-Of... or Just Plain Impersonation?
-
Validation: Foundry Guardrails 15:15
-
In Practice: Foundry Guardrails 3:43
-
Validation: Middleware Architecture 13:04
-
Who Guardrails the Guardrail?
-
In Practice: Coding Middleware with the MAF 2:52
-
Deployment Strategies 16:25
-
Rollbacks, Replay and Restore-Safe Design 18:11
-
The Three R's (No, Not Those): Retry, Replay and Restore
-
The Retry Ladder
-
Quotas, Allocation and Chargeback 14:05
-
RAG Security in Multitenant 15:03
-
Contains: From Dev to Deploy (V2) 14:20
-
Disabling Public Access ≠ Being Reachable Privately!
-
In Practice: The Azure Container Registry 1:54
-
Service-Managed vs. Local Agents 18:40
-
In Practice: Foundry Hosted Agents 2:29
-
Module Outro 3:31
-
Domain 4 Quiz (10 Questions)
-
Key Distinctions - Domain 4
-
Role Play: Interviewing for Multi-Agent Security Lead
-
Graded Lab: Fix the Secure and Deploy Settings
Chương 8 Reviewing the AI-500 Certification 6 bài · 50 phút
-
Module Intro 1:58
-
Cross-Domain Patterns 11:31
-
Key Traps to Avoid 15:52
-
The 10 Takeaways, Revisited 18:40
-
Module Outro 1:49
-
Graded Lab: Set the Review Judgment Dials
Chương 9 The Triangle Prep: Preparing for the Exam's Traps 7 bài
-
The Triangle Prep Framework
-
Group 1: Language Traps
-
Trap Batch: Language Traps (10 Questions)
-
Group 2: Content Traps
-
Trap Batch: Content Traps (10 Questions)
-
Group 3: Case Study Extraction
-
Trap Batch: Case Study Extraction (10 Questions)
Chương 10 Course Conclusion 1 bài · 4 phút
-
Course Conclusion 4:04
Requirment
-
A basic understanding of Azure AI, LLMs, and agentic systems is recommended, although foundational concepts are recapped.
-
AI-103 is not required to take this course or sit AI-500, but the AI-103 certification is required when the AI-500 credential is awarded.
Outcomes
-
You will learn how LLM agents reason, use tools, manage context and memory, coordinate workflows, and stop safely.
-
You will learn to architect secure, stateful multi-agent solutions using orchestration patterns, trust boundaries, identity, and RBAC.
-
You will learn to develop Azure multi-agent systems with RAG, MCP, A2A, Cosmos DB, Foundry SDK, Agent Framework, and LangGraph.
-
You will learn to evaluate agents with Foundry evaluators, golden tests, LLM-as-judge, synthetic data, red teaming, and human review.
-
You will learn to monitor production workflows through telemetry, tracing, checkpoints, replay, failure classification, and recovery controls.
-
You will learn to secure and deploy agentic systems using guardrails, middleware, tenant isolation, canary releases, and rollback.
-
You will learn the AI-500 domains, exam process, recurring architecture patterns, common traps, and study priorities.
0 Student
11 Courses
0 review