🤖 Azure AI Foundry + Hosted Agents
Track Type: Skill Track — Deep Technical
Target Audience: AI Solution Architects, Cloud Architects, Principal Engineers
Prerequisites: AZ-900 or equivalent Azure exposure; Python or PowerShell comfort
What You'll Build
Production-grade AI architectures grounded in real enterprise constraints. Every challenge starts from a customer scenario — the kind you'll encounter in actual customer engagements.
Platform Context (2025)
Microsoft rebranded Azure AI Studio → Azure AI Foundry → Microsoft Foundry (2025). The platform unifies:
| Component | Role |
|---|---|
Foundry Portal (ai.azure.com) | Unified project hub |
| Foundry Resource | Single account + projects (replaces Hub + AOAI + AI Services) |
| Agent Service | Three agent types: Prompt, Workflow, Hosted |
| Model Catalog | 1,900+ models from Microsoft, OpenAI, Meta, Mistral, Hugging Face |
| Tool Ecosystem | 1,400+ MCP connectors, Azure Search, Logic Apps, Functions |
| Evaluation SDK | Groundedness, coherence, fluency, safety scoring |
| Observability | OpenTelemetry traces, dashboards, Azure Monitor integration |
The Three Agent Types
┌─────────────────────────────────────────────────────────────┐
│ Azure AI Agent Service │
├──────────────┬──────────────────┬───────────────────────────┤
│ Prompt Agent │ Workflow Agent │ Hosted Agent │
│ (No-code) │ (YAML/Visual) │ (Container/Code) │
│ │ [Preview] │ [Preview] │
├──────────────┼──────────────────┼───────────────────────────┤
│ Chat UI │ Multi-step flows │ Full code control │
│ Simple Q&A │ Conditional DAGs │ BYO framework (LangGraph) │
│ 5-min deploy │ Visual designer │ Micro-VM isolation │
│ │ │ <1s cold start │
│ │ │ $0 idle cost │
└──────────────┴──────────────────┴───────────────────────────┘
Hosted Agents — Architecture Deep Dive
Hosted Agents are the enterprise-grade, code-first option. Key facts every architect must know:
- Isolation: Each agent runs in a Micro-VM (gVisor-based) — process-level isolation, not shared container
- Lifecycle: Cold start < 1 second; $0 idle cost; scales to zero automatically
- Frameworks: LangGraph, Azure Agent Framework, Semantic Kernel — bring any Python agent
- Network: Supports BYO VNet (Standard mode) — private endpoints, NSG integration
- Identity: Every agent gets an Entra Agent ID (managed identity) — no passwords, full RBAC
- API Version: Build on Responses API (v2) — the classic Assistants API (v1) retires March 2027
- SDK:
azure-ai-projects2.x — unified client, single project endpoint
Deployment Modes: Critical Enterprise Decision
| Mode | Storage | Networking | Use Case |
|---|---|---|---|
| Basic | Microsoft manages | Public | Dev/test only |
| Standard | Customer BYO (Storage + Key Vault + Search) | Private endpoints available | All enterprise workloads |
Always Use Standard Mode for Enterprise
Basic mode gives Microsoft access to your conversation data and artifacts. Any customer with data residency, HIPAA, PCI, or EU AI Act requirements must use Standard mode.
Learning Path (14 Modules)
| # | Module | Type | Time |
|---|---|---|---|
| 1 | Platform Architecture & Resource Model | Concept | 45 min |
| 2 | Agent Service Fundamentals | Lab | 60 min |
| 3 | First Hosted Agent (LangGraph) | Lab | 90 min |
| 4 | VNet Isolation & Private Endpoints | Lab | 90 min |
| 5 | Entra Agent ID & RBAC | Lab | 60 min |
| 6 | Evaluation SDK & Groundedness | Lab | 90 min |
| 7 | Content Safety Integration | Lab | 60 min |
| 8 | Multi-Agent Orchestration | Lab | 120 min |
| 9 | MCP Tools & External Connectors | Lab | 90 min |
| 10 | OpenTelemetry & Observability | Lab | 90 min |
| 11 | Standard Mode Enterprise Setup | Lab | 120 min |
| 12 | Production Troubleshooting | Lab | 90 min |
| 13 | Agent Publishing (Copilot Studio / Teams) | Lab | 60 min |
| 14 | Cost Optimization & Scaling | Concept | 45 min |
Challenges (Scenario-Based Labs)
All labs are based on real enterprise scenarios:
| Challenge | Scenario | Domain |
|---|---|---|
| 1 — BYO VNet Healthcare | Patient data can't leave hospital VNet | Healthcare / Data Residency |
| 2 — Hallucination in Production | Agent returns wrong answers 20% of the time | Evaluation / Quality |
| 3 — Entra Agent ID Setup | Standard mode with managed identity + RBAC | Security / Identity |
| 4 — Multi-Agent Claims System | Bank security team reviewing AI architecture | Financial / Multi-Agent |
| 5 — Hosted Agent at Scale | Works locally, fails at 500 concurrent users | Production / Troubleshooting |