Welcome to the first practice set for AI-103, the brand-new Azure AI Apps and Agents Developer Associate exam. This post focuses on building Azure AI Foundry agents. This topic has dominated cloud news since Microsoft Build 2026. That release brought hosted agent runtimes, Foundry IQ grounding, and 1,400+ MCP-connected tools to general availability. For exam candidates, this is the largest domain on the blueprint at 30-35% of scored items. Therefore, it deserves your first real study block.
AI-103 replaced AI-102 after the older exam retired on June 30, 2026. If you were mid-study on AI-102, most of your knowledge transfers, however the emphasis has shifted decisively toward generative AI and agentic workflows. Therefore, expect scenario questions that hinge on which Foundry primitive to use, not just what each service does.
This is Questions 1-10, a mix of easy warm-ups, medium design choices, and a few hard troubleshooting scenarios. Reveal the answer only after you commit to a choice — the explanations teach the “why,” which is where the exam separates confident engineers from guessers.
Question 1: Choosing the Right Agent Platform
Contoso is starting its first production agent project. The engineering team writes Python daily. They need fine-grained control over model selection and RAG pipelines. In addition, they must ship a custom evaluation harness that plugs into their existing OpenTelemetry stack. Which Microsoft platform best fits these requirements?
A) Microsoft Copilot Studio, because it offers the fastest time to production for any agent scenario.
B) Azure AI Foundry Agent Service, because it is a pro-code platform that exposes the full agent lifecycle and native tracing.
C) Power Automate flows with an HTTP action calling Azure OpenAI directly.
D) Microsoft 365 Agents SDK with the built-in low-code designer.
👁 Reveal Answer
Correct Answer: B
Explanation: Azure AI Foundry Agent Service is Microsoft’s pro-code agent platform. It gives engineering teams direct control over model choice, RAG pipelines, evaluation, and OpenTelemetry-based observability. Copilot Studio (A) is optimized for low-code business makers, not for a team that wants custom evaluation and deep model tuning. Power Automate (C) can call models but lacks native agent primitives such as tools, threads, and grounding. The Microsoft 365 Agents SDK (D) focuses on Teams and Copilot channel delivery rather than pro-code Azure AI Foundry agents.
Question 2: Grounding an Agent in Enterprise Data
The engineering team at Fabrikam needs its support agent to answer questions using content from SharePoint sites, a Microsoft Fabric lakehouse, and public Bing results. Leadership wants a single grounded knowledge source that respects existing Microsoft 365 permissions. Which capability should the team enable first?
A) Custom-built ingestion pipelines pushing every document into Azure Cognitive Search.
B) Foundry IQ, which unifies SharePoint, Fabric, and Bing as grounding sources for Azure AI Foundry agents.
C) Azure Data Factory copying files nightly into a blob container.
D) Direct Graph API calls embedded in a custom tool.
👁 Reveal Answer
Correct Answer: B
Explanation: Foundry IQ is Microsoft’s managed grounding layer. It natively connects to SharePoint, Fabric, and Bing while honoring existing permissions. As a result, no custom ingestion is needed. Building your own pipeline into Azure AI Search (A) works. However, it re-implements what Foundry IQ ships out of the box. It also forces you to reconcile permissions manually. Azure Data Factory (C) can move data but adds nightly latency and does not enforce Microsoft 365 access controls. A custom Graph tool (D) can serve narrow use cases, however it scales poorly across three sources and skips built-in trust-boundary handling.
Question 3: Delegating Work to Specialist Agents
Woodgrove Bank has one orchestrator agent that handles customer chat plus three specialist agents for fraud checks, credit inquiries, and account changes. The team wants the orchestrator to hand off tasks in natural language without hand-coded routing. Which pattern should they implement?
A) Have the orchestrator call each specialist through a REST endpoint and parse text responses.
B) Register each specialist as a connected agent so the orchestrator can invoke it as a tool over A2A.
C) Duplicate every specialist’s instructions inside the orchestrator’s system prompt.
D) Run each specialist inside a separate Azure Function triggered by a service bus queue.
👁 Reveal Answer
Correct Answer: B
Explanation: Connected agents let one Azure AI Foundry agent register another as a callable tool. The Agent-to-Agent (A2A) protocol then handles delegation in natural language. Therefore, no custom routing code is needed. Manual REST calls (A) reintroduce the exact plumbing A2A eliminates. Copying specialist prompts into the orchestrator (C) creates a bloated context, brittle updates, and no separation of concerns. Service bus queues (D) are appropriate for asynchronous work but overkill for synchronous natural-language handoff between Azure AI Foundry agents.
Question 4: Long-Running, Stateful Approval Flows
Your team is modeling a mortgage approval process that spans several days, requires human sign-off, and must retry a failed credit check up to three times. The design must survive process restarts. Which Foundry construct is the best fit?
A) A single connected agents graph called once per request.
B) Multi-agent workflows, which add a stateful layer for context, retries, compensation, and long-running steps.
C) A prompt-flow evaluation run scheduled every hour.
D) Chained Logic App triggers with no shared state.
👁 Reveal Answer
Correct Answer: B
Explanation: Multi-agent workflows in Foundry provide durable state, checkpointing, compensation, and step-level retries — exactly what a multi-day approval needs. Connected agents (A) are ideal for synchronous delegation, however they do not persist state across days or handle compensation logic. Prompt-flow evaluations (C) measure quality; they are not an execution runtime. Chained Logic Apps (D) can be stitched together but lack a native shared context, so you would end up rebuilding what workflows already provide.
Question 5: Adding External Tools to an Agent
Northwind Traders wants its agent to call a public weather API, an internal inventory service, and GitHub. The company prefers standard, portable tool definitions rather than one-off wrappers. Which approach best supports Azure AI Foundry agents at scale?
A) Hand-code a Python function for each API and register it as a function tool per agent.
B) Connect the agent to MCP-enabled tools, since Foundry supports 1,400+ Model Context Protocol integrations.
C) Build custom Azure API Management policies to intercept model output.
D) Store API responses in a SharePoint list and let Foundry IQ read them.
👁 Reveal Answer
Correct Answer: B
Explanation: Model Context Protocol (MCP) is the emerging standard for exposing tools to LLMs. Foundry’s Toolboxes surface 1,400+ MCP-enabled integrations. As a result, any agent can reuse them. Custom Python functions (A) are still useful for narrow logic, however they do not scale across many APIs the way MCP does. API Management (C) is for gateway policies, not tool discovery. SharePoint lists (D) can store data but are the wrong abstraction for calling live APIs.
Question 6: Deploying an Existing LangGraph Agent
Tailwind Traders already runs a LangGraph agent packaged as a container image. Ops wants a managed runtime with autoscaling, session isolation, and zero cost when idle — without rewriting the agent into a new framework. Which target should they deploy to?
A) Azure App Service on a Basic plan.
B) An Azure Kubernetes Service cluster the team operates themselves.
C) Hosted agents in Foundry Agent Service, which accept containers and provide session-isolated runtime with autoscaling.
D) Azure Functions Premium plan with a custom handler.
👁 Reveal Answer
Correct Answer: C
Explanation: Hosted agents in Foundry Agent Service are framework-agnostic. LangGraph, OpenAI Agents SDK, Microsoft Agent Framework, or custom code can all be pushed as container images. As a result, you get session-isolated runtime, autoscaling, and zero idle cost. App Service Basic (A) has no scale-to-zero and lacks per-session isolation. Self-managed AKS (B) works but shifts operational burden back onto the team. Functions Premium (D) supports containers but is not optimized for long-lived agent sessions with persistent state.
Question 7: Tracing and Evaluating a Production Agent
During a rollout at Litware, an agent occasionally returns off-topic answers. The team must trace every model call, tool invocation, and sub-agent hop back to the failing turn, and correlate that trace with quality evaluations. Which Foundry capability should they configure?
A) Application Insights alone, correlating traces manually by request ID.
B) Foundry’s built-in OpenTelemetry pipeline, which links every model call, tool call, sub-agent hop, and evaluation into one trace.
C) A custom Log Analytics workspace with regex-based parsing of agent logs.
D) Enable content filters at the model level.
👁 Reveal Answer
Correct Answer: B
Explanation: Foundry pipes every model call, tool invocation, sub-agent hop, and handoff through one OpenTelemetry pipeline. Evaluations link directly back to the trace. Therefore, a single click takes you from a failing turn to the offending step. Application Insights (A) can ingest OTel data but requires manual correlation. Custom regex on logs (C) is fragile and does not join traces to evaluation results. Content filters (D) reduce specific harm categories but do nothing for troubleshooting off-topic behavior.
Question 8: Function Calling vs Logic Apps as Tools
Adventure Works has an agent that must both create Salesforce cases (through an existing Logic App) and compute a proprietary risk score in Python. The team wants to expose both capabilities to the agent using the least custom code. Which combination is correct?
A) Wrap both in Azure Functions and call them through function tools.
B) Register the Logic App as a Logic Apps tool and expose the Python risk score as a function tool via function calling.
C) Rewrite the Logic App as Python code so both can be function tools.
D) Use Foundry IQ to invoke both connectors.
👁 Reveal Answer
Correct Answer: B
Explanation: Foundry’s Agent Service natively exposes Logic Apps as tools. Therefore, the existing Salesforce flow needs no rewrite. In addition, function calling is the right primitive for the Python risk score. Wrapping the Logic App in a Function (A) throws away working orchestration. Rewriting the Logic App (C) is wasted effort. Foundry IQ (D) is a grounding surface for knowledge sources, not a general connector runtime.
Question 9: Filtering Harmful Content in Agent Output
A retail startup is preparing its first customer-facing Azure AI Foundry agent for launch. Compliance requires blocking self-harm and hate content in both prompts and completions and logging every triggered filter. Which configuration meets the requirement most directly?
A) Rely on the base model’s built-in safety training and monitor complaints.
B) Configure Azure AI Content Safety content filters on the deployment with prompt and completion detection enabled, and route filter events into the OpenTelemetry trace.
C) Add a custom Python regex tool that scans outputs before returning them.
D) Enable Azure Defender for Cloud on the resource group.
👁 Reveal Answer
Correct Answer: B
Explanation: Azure AI Content Safety filters run at the model deployment layer. They cover both prompt and completion sides. In addition, they emit events that flow into Foundry’s OpenTelemetry pipeline for auditing. Trusting model training alone (A) fails the compliance requirement because it provides no per-event log. Regex-based scanning (C) is brittle and misses paraphrased harms. Defender for Cloud (D) protects Azure infrastructure but does not evaluate LLM content.
Question 10: Choosing a Retrieval Strategy for a RAG Agent
Continuing this review, an insurance carrier is grounding an Azure AI Foundry agent in policy PDFs. Users ask both keyword-heavy questions (“policy 47B”) and vague semantic ones (“what if my roof leaks?”). The team must maximize recall across both styles while keeping latency low. Which retrieval strategy should they choose?
A) Pure keyword search (BM25) in Azure AI Search.
B) Pure vector search on embeddings.
C) Hybrid search combining BM25 and vector search, with semantic ranking on top.
D) A nightly full-text export to Azure SQL and a LIKE query.
👁 Reveal Answer
Correct Answer: C
Explanation: Hybrid search combines lexical (BM25) and semantic (vector) recall. Azure AI Search’s semantic ranker then reorders the top candidates. Specifically, Microsoft recommends this pattern for mixed keyword-and-semantic workloads. Pure BM25 (A) misses paraphrased questions. Pure vector search (B) can under-perform on exact identifiers like “policy 47B.” A nightly SQL LIKE query (D) has stale data and no relevance ranking. As a result, it cripples grounding quality for Azure AI Foundry agents.
Study Tips for Building Azure AI Foundry Agents
Before you sit AI-103, spend a full study block inside the Foundry portal — reading about agents is not the same as building one. Use the following tips to focus your time.
- Learn the primitives by name. The exam expects you to distinguish an agent, a tool, a connected agent, a workflow, and a thread. If you can draw the diagram, you can pass the questions.
- Practice picking the right platform. Copilot Studio, Foundry Agent Service, and the Microsoft 365 Agents SDK all build agents. For example, Copilot Studio suits low-code makers, while Foundry Agent Service targets pro-code engineers.
- Trace one real request end to end. Deploy a small agent, watch the OpenTelemetry trace, and correlate a model call to its tool invocation. Specifically, this cements how evaluation ties back to spans.
- Study grounding beyond one source. Foundry IQ, Azure AI Search hybrid retrieval, and MCP tools each play a distinct role in production Azure AI Foundry agents.
- Know your responsible AI defaults. Content Safety filters, the Agent Control Specification (ACS), and Rubric evaluation are all fair game. Therefore, review Microsoft’s trust tooling before test day.
Keep Practicing Your Azure AI Foundry Agents Skills
You just finished the first ten questions on Azure AI Foundry agents for AI-103. Save the ones you missed. Re-read the explanation the next morning. Then revisit the linked Microsoft Learn material until you can teach the pattern to a teammate.
For the official blueprint, work through the Microsoft Learn AI-102/AI-103 study guide, which still maps cleanly to the current agent objectives. In addition, ground your foundation with our AI-102 Questions 1-10 on Generative AI Solutions and reinforce retrieval concepts with AI-102 Questions 31-40 on Knowledge Mining. If you also design broader infrastructure, our AZ-305 Questions 41-50 mixed review pairs well with agent architecture thinking.
Questions 11-20 will go deeper into evaluation, tracing, and responsible AI for Azure AI Foundry agents. See you tomorrow — and good luck.
