Microsoft's New Azure AI Exam Covers Foundry's Model Catalog, Router, and Agents
This survey of Azure AI Foundry maps its core components to the new exam blueprint, covering the model catalog, router, and agent framework. It also explores deployment options, API differences, and cost considerations, with examples from real usage.
Listen to this dispatch
Narrated by an AI-generated voice.
The Microsoft Azure certification track is being updated as the new exam, "Develop AI apps and agents on Azure," supersedes AI-102. A survey of Azure AI Foundry's current state is organized around the platform's core components, each mapped to a section of the exam blueprint: a model catalog of more than 10,000 entries, a routing layer that picks models on the fly, and an agent framework with guardrails on by default.
Foundry's Structure and Model Catalog
The survey opens by establishing Foundry's structure—a resource containing isolated projects, each holding models, agents, tools, and knowledge. It notes that 102 models are sold directly by Azure, while the "all models" view exceeds 10,000, with roughly 11,000 from Hugging Face. Open-weight models can be downloaded and run anywhere with sufficient compute. They offer privacy, fine-tuning freedom, and permanence: "I have the weights. I could run this thing forever." However, open weight does not imply small; some models have trillions of parameters.
Deployment Options
Deployment comes in two billing models. Serverless provides a private endpoint and charges per token. Managed compute reserves dedicated CPU or GPU instances, bills by the hour and instance count, and auto-scales with traffic. A third path is Fireworks AI, a compilation and runtime engine for low-latency execution of open-weight models, billed on tokens consumed.
Model Router
The exam expects understanding of the model router, a newer piece that lets apps talk to a router configured with a set of models. A small language model decides which model handles each request. The benefit is cost optimization and automatic access to new models. The concern is detecting regression if models change underneath an application. The answer is the evaluation catalog: automated test sets, often judged by another language model, run against production traces. Evaluations receive emphasis because generative AI is not deterministic.
API Distinctions
The demo code highlights a distinction: the Responses API keeps conversation state server-side, so a second request like "what about France" understands context. The Chat Completions API stores nothing server-side, requiring the client to resend the full history each turn, so payloads grow.
Agent Framework
The agent framework supports multi-agent patterns (concurrent, sequential, group chat) and agent memory that persists preferences across sessions. Guardrails default to blocking jailbreaks, with configurable filters for indirect prompt injection, hate, self-harm, violence, and PII leakage. Knowledge comes through Foundry IQ, built on what was Azure AI search, and returns citations. Hosted agents run as containers on managed endpoints and can be published to Teams, the web, or M365 Copilot.
Pre-Built Services and Costs
Pre-built services round out the catalog: language detection (returning 100% confidence in demos), PII redaction, speech, and Azure Content Understanding, which converts documents, audio, and video into structured data. The advice is to use them rather than rebuilding. Costs reinforce that: four seconds of generated video ran $0.40, and total usage over seven days was $0.10.