Version : 1.0 · Date : 2026-10-22
Cours complet pour l'examen AI-500 : Designing and Implementing Multi-Agent AI Solutions (certification Microsoft Certified: Multi-Agent AI Solutions Expert, prérequis : la certification AI-103). Un chapitre par bloc de compétences, avec schémas draw.io, Mermaid, code Python (Microsoft Agent Framework, Foundry SDK), labs, pièges d'examen et scénarios corrigés. Toutes les informations viennent de Microsoft Learn (documentation et modules du cours AI-500T00-A) et du dépôt officiel Agent Framework, lus les 21 et 22/10/2026. Les noms de services, fonctionnalités, rôles et paramètres sont en anglais.
L'examen est en bêta et le guide n'a pas de date de version (vérifié inchangé le 22/10/2026). Foundry évolue vite et beaucoup de fonctionnalités sont en préversion : chaque écart est signalé par la réponse attendue à l'examen, puis l'état actuel. Relire le guide AI-500 juste avant l'examen.
Les domaines de l'examen#
| Domaine (intitulé officiel, en anglais) | Poids | Chapitres |
|---|---|---|
| Architect multi-agent solutions | 15 à 20 % | 01, 02 |
| Develop multi-agent solutions in Azure | 30 à 35 % | 03, 04, 05 |
| Evaluate, optimize, and monitor multi-agent solutions | 20 à 25 % | 06 |
| Secure, govern, and deploy multi-agent solutions | 20 à 25 % | 07 |
Score de réussite : 700 sur 1000. Le développement pèse un tiers de l'examen : chapitres 03 à 05 en priorité. Le cours suppose AI-103 acquis (Foundry, agents, RAG) et renvoie à AI-300 (MLOps, évaluation) et Claude Architecte (motifs d'agents) au lieu de les répéter.
Les chapitres#
Chaque chapitre suit le même plan : ce que l'examen demande (texte officiel), concepts, schémas, mise en œuvre, tableaux comparatifs, labs conseillés, pièges d'examen, scénarios corrigés, pour aller plus loin, fiche à retenir et références.
Couverture des objectifs de l'examen#
Les 55 compétences du guide d'étude, chacune reliée à la section qui la traite (vérification du 22/10/2026).
Architect multi-agent solutions (15 à 20 %)#
| Ch. | Compétence mesurée (guide officiel, en anglais) | Section |
|---|---|---|
| 01 | Decompose goals and objectives into workflows, agents, and tools | 1 |
| 01 | Design workflows, including agents, subagents, control loops, and human-in-the-loop | 2 |
| 01 | Specify agent personas, scopes, boundaries, autonomy levels, and behavioral guidelines | 3 |
| 01 | Specify tool scopes, permission boundaries, and authentication methods | 4 |
| 01 | Specify appropriate protocols for communication, including between multiple agents and between agents and other components | 5 |
| 01 | Design controls and functionality to support human-AI experience (HAX) and to adhere to responsible AI principles | 6 |
| 01 | Design short-term and long-term memory architectures, including context sharing | 7 |
| 01 | Match task demands to model family capacities | 8 |
| 02 | Specify integration components to coordinate agent-to-agent communication, agent-to-tool interactions, and agent-to-knowledge-source routing across a multi-agent production system | 2 |
| 02 | Specify components for Zero Trust multi-agent solutions, including per-agent identity scoping, lateral movement prevention, and compliance control mapping for regulated deployments | 3 |
| 02 | Specify multi-tier state persistence components that address session state, shared team state, and long-term semantic memory with lifecycle and tenant-isolation policies | 4 |
| 02 | Specify compute components for multi-agent solutions that meet scalability, reliability, operational efficiency, security, and cost requirements | 5 |
| 02 | Specify observability components, including cross-service trace correlation, structured logging of agent reasoning paths, and agent replay capture for reproducible debugging | 6 |
| 02 | Specify monitoring components, including cross-agent coordination tracking, behavioral drift and quality regression detection, and automated remediation for agent failure patterns | 7 |
| 02 | Specify developer tools and environment to support the Software Development Lifecycle (SDLC), including dev containers, Visual Studio Code extensions, command-line interface (CLI), dependency management, and AI instructions | 8 |
Develop multi-agent solutions in Azure (30 à 35 %)#
| Ch. | Compétence mesurée (guide officiel, en anglais) | Section |
|---|---|---|
| 03 | Design and implement advanced prompt engineering techniques, including examples, dynamic context injection, defensive guidelines, and prompt lifecycle management | 2 |
| 03 | Design and implement context-aware multi-agent behaviors | 3 |
| 03 | Design and implement a strategy for fine-tuning agents and models, including data and frequency | 4 |
| 03 | Design and implement context management in single agents and between agents, including context accumulation, retrieval, injection, and compaction | 5 |
| 03 | Design and implement single-agent and multi-agent memory strategies, including security, compliance, lifecycle, storage, and session management | 6 |
| 03 | Design and implement a multi-agent RAG architecture, including chunking, embedding quality, and retrieval precision | 7 |
| 03 | Design and implement knowledge integration for multi-agent consumption, including search, RAG, MCP-available sources, and semantic search | 8 |
| 04 | Integrate external resources, including function calling, specified tool use, and dynamic tool use | 2 |
| 04 | Design and build MCP servers and clients, including Azure Functions, Azure Logic Apps, and Azure API Management | 3, 4 |
| 04 | Design tool error-handling and fallback mechanisms | 5 |
| 04 | Implement tool result validation and quality checks | 6 |
| 05 | Implement multi-agent orchestration patterns, including hub-and-spoke, sequential, parallel, peer-to-peer, and orchestrator-subagent | 1 |
| 05 | Implement human-in-the-loop processes, including approval workflows, overrides, and edge case handling | 2 |
| 05 | Design and implement a caching strategy, including prompt caching, semantic caching, and response caching | 3 |
| 05 | Implement scaling and control of multi-agent spawning, batching, and concurrent execution | 4 |
| 05 | Integrate existing agents into a multi-agent solution securely by using Agent2Agent (A2A) and/or MCP | 5 |
| 05 | Implement multi-agent orchestration frameworks, including Agent Framework, LangChain, and LangGraph | 6 |
| 05 | Implement advanced multi-agent capabilities by using Hugging Face Transformers | 7 |
| 05 | Design middleware for reusable capabilities across the solution, including logging, authorization, and exception handling | 8 |
Evaluate, optimize, and monitor multi-agent solutions (20 à 25 %)#
| Ch. | Compétence mesurée (guide officiel, en anglais) | Section |
|---|---|---|
| 06 | Design and implement a human review process to evaluate solutions in Foundry | 2 |
| 06 | Design and implement evaluations for memory, knowledge, tools, and prompts | 3 |
| 06 | Optimize task duration, including parallelism and rate limits | 4 |
| 06 | Diagnose context window issues, including sliding-window amnesia, summary drift, vector-only recall, and entity continuity issues | 5 |
| 06 | Design and implement continuous improvement, including large language model–as-a-judge (LLM-as-a-judge) frameworks, synthetic data generation, semantic optimization loops, and user feedback loops | 6 |
| 06 | Design a strategy for monitoring reliability, including agent health, workflow failures, trace correlation, drift detection, quality regression, and remediation | 7 |
| 06 | Design a strategy for monitoring solution and platform availability, performance, reliability, and service-level agreement adherence | 8 |
| 06 | Design and implement a strategy for optimizing token usage, including token limits, loop controls, and tool calls | 9 |
| 06 | Design a strategy for monitoring and managing cost, including usage, quotas, allocations, and chargebacks | 10 |
| 06 | Implement tracing in Foundry, including tokens, prompts, correlation IDs, alerting, and execution tracking | 11 |
Secure, govern, and deploy multi-agent solutions (20 à 25 %)#
| Ch. | Compétence mesurée (guide officiel, en anglais) | Section |
|---|---|---|
| 07 | Specify and implement resource access, including identity-based access, network boundaries, access control policies, and role-based access control (RBAC) | 2 |
| 07 | Design and implement authentication flows in a multi-agent solution, including user impersonation, on-behalf-of, API keys, and OAuth 2.0 | 3 |
| 07 | Design the secrets management and retrieval approach for multi-agent solutions by using Azure Key Vault, including secrets, certificates, key rotation, role-based access to secrets, and encryption | 4 |
| 07 | Implement shift-left security principles, including the AI Red Teaming Agent in Foundry | 5 |
| 07 | Design and implement a multi-intervention guardrail strategy covering user inputs, tool calls, tool responses, and outputs | 6 |
| 07 | Design and implement custom guardrails for domain-specific constraints | 7 |
| 07 | Design guardrail testing and validation, including using synthetic data | 8 |
| 07 | Choose a release methodology, including Development-Test-Acceptance-Production (DTAP), blue/green, and canary | 9 |
| 07 | Design and implement multi-environment release strategies, including rollback strategies, release management, and rollout | 10 |
| 07 | Design and implement a strategy for testing, including unit tests, regression tests, integration tests, and automated evaluations | 11 |
| 07 | Specify continuous integration and continuous deployment (CI/CD) requirements for integration and release, including infrastructure-as-code deployments and testing | 12 |
Organisation du dossier#
AI-500/
├── README.md ce plan de révision
├── 01-…md à 07-…md un fichier par bloc de compétences
├── images/ les schémas exportés en PNG (affichés dans les chapitres)
├── DRAWIO/ les schémas sources, modifiables dans draw.io
└── outils/ les scripts qui génèrent les schémas (moteur de GenDoc)
Régénérer les schémas d'un chapitre : python outils/schemas_NN.py [noms…] depuis ce dossier (GenDoc doit être dans ..\..\GENERATEUR-PPT).
Méthode de révision conseillée#
- Revoir AI-103 si besoin, puis lire 01 et 02 (architecture) avant 03 à 05 (développement).
- Faire les labs conseillés dans un projet Foundry de test, avec un budget et des quotas bas, puis supprimer le resource group.
- Répondre aux scénarios sans regarder la réponse (pas d'évaluation d'entraînement officielle annoncée).
- La veille de l'examen : relire les fiches à retenir et les tableaux de pièges.