Domain 3: Implement an Agentic Solution (5-10%)
This is a new domain for 2025 focusing on autonomous AI agents that can use tools and reason through multi-step tasks. It carries the lowest weight on the exam (5–10%), but the concepts are distinct enough to trip up candidates who confuse agent patterns with standard LLM calls.
3.1 Azure AI Agent Service
A fully managed service in Microsoft AI Foundry for building, deploying, and scaling AI agents. It abstracts the orchestration loop so you focus on configuring tools and instructions rather than managing state.
Reasoning Loop
An agent doesn't generate a single response — it iterates:
- Receive user goal
- Plan — decide if a tool is needed
- Execute — call the tool (code, search, function)
- Observe — process the tool result
- Repeat or Respond — loop until the goal is met or budget is exhausted
Key Exam Concept
The phrase "autonomous multi-step task" in a question → answer is AI Agent Service, not a plain Prompt Flow run or a single LLM call.
Core Agent Components
| Component | What It Is | Exam Signal |
|---|---|---|
| Instructions | System-level prompt that defines the agent's persona and constraints | "agent must always respond in English" |
| Thread | A conversation session — stores the message history | "maintain conversation state across turns" |
| Run | A single execution of the reasoning loop on a Thread | "trigger the agent to process a message" |
| Tool | A capability the agent can invoke (built-in or custom function) | "agent needs to call an API" |
Built-in Tools
| Tool | Purpose | When to Use |
|---|---|---|
| Code Interpreter | Writes and executes Python in a secure sandbox | Math, data analysis, file conversion, chart generation |
| File Search | Retrieves information from uploaded documents via vector search | "agent searches uploaded PDFs", "agent answers from your files" |
| Function Calling | Calls a custom function/API you define | Real-time data, internal systems, custom logic |
| Azure AI Search | Uses an external AI Search index as a grounding source | "agent searches your indexed knowledge base" |
Code Interpreter vs File Search
Code Interpreter → agent writes and executes code to solve a problem (e.g., calculate statistics, generate a chart). File Search → agent retrieves information from documents without executing code.
The exam distinguishes these with phrases like "solve math problems" (Code Interpreter) vs "answer from uploaded documents" (File Search).
Configuration Reference
# Create an agent with tools
agent = client.agents.create_agent(
model="gpt-4o",
name="data-analyst",
instructions="You are a data analyst. Use code interpreter to analyze data.",
tools=[{"type": "code_interpreter"}]
)
# Create a thread and run
thread = client.agents.create_thread()
client.agents.create_message(thread_id=thread.id, role="user", content="Summarize this CSV")
run = client.agents.create_and_process_run(thread_id=thread.id, agent_id=agent.id)Async Pattern
Agent runs are asynchronous. create_and_process_run() polls until completion. For manual polling, use create_run() then GET the run status until completed or failed.
3.2 Microsoft Agent Framework
The Agent Framework provides patterns and components for building sophisticated multi-agent applications. It integrates with Semantic Kernel for orchestration logic.
Key Components
| Component | Purpose |
|---|---|
| Planner | Breaks a high-level goal into a sequence of tool calls / sub-tasks |
| Persona | Defines the agent's name, description, and behavioral constraints |
| Memory | Persistent storage across agent turns (conversation history, retrieved facts) |
| Kernel | The Semantic Kernel engine that connects the model to tools and memory |
Multi-Agent Orchestration Patterns
When a single agent isn't enough, multiple specialized agents can collaborate:
| Pattern | Structure | Best For |
|---|---|---|
| Hierarchical | A "Manager" agent delegates sub-tasks to "Worker" agents | Complex workflows where tasks can be decomposed |
| Sequential | Agents form a pipeline — each passes output to the next | Document processing, staged analysis |
| Group Chat / Joint | Agents discuss and critique each other's responses | Research tasks requiring multiple perspectives |
Exam Trigger
"Multiple agents collaborating" → answer involves multi-agent orchestration (not a single agent with multiple tools).
3.3 Testing, Monitoring, and Deployment
Constraints and Safety
- Max turns: Set a budget on the reasoning loop to prevent infinite tool-calling cycles.
- Content filters: Applied at the model level via Azure OpenAI content filtering settings.
- Grounding checks: Use Prompt Shields to detect prompt injection attempts within tool outputs.
Evaluation
- Foundry Tracing: Captures every step of the reasoning loop — tool calls, inputs, outputs, latency. Essential for debugging "why did the agent call that tool?"
- Evaluation Flows: Run the agent against a dataset of golden Q&A pairs; measure groundedness, relevance, and task completion rate.
Deployment
Agents are deployed as scalable endpoints within an AI Foundry Project. They inherit the project's security, connections, and compute settings from the parent Hub.
Scope Trap
The exam may ask where agents are deployed. Agents live in a Project (your workspace), not directly in a Hub. The Hub provides shared resources; the Project is where you build and deploy.