Track D: Production (M9-10) — Chat interfaces
& Azure Functions
Reference Architecture & Pacing — Full system
integration & demo pacing
Theory 1
What are Agents?
An agent is an autonomous software component that uses an LLM as
its central reasoning engine to perceive its environment, make
decisions, call APIs, and iteratively reason toward completing
complex goals.
🧠LLM CorePlan · Reason · Decide
📋Planning
Decomposes complex goals into ordered, logical steps and
iteratively refines strategy
🛠️Tools
Calls APIs, reads files, runs health checks, and executes
Python functions on demand
💾Memory
Session history for context continuity and long-term
persistent data via ContextProvider
👁️Perception
Observes environment, tool outputs, and external signals to
adapt reasoning in real-time
Theory 2
What is Microsoft Agent Framework?
An open-source framework for Python and .NET
that helps teams build production-grade AI systems using autonomous
agents and explicit workflows.
Two Core Capabilities: Use
Agents for conversational and tool-driven tasks,
or Workflows for deterministic multi-step orchestration.
Foundational Building Blocks: Model clients,
agent sessions, context providers, middleware, and MCP clients
for robust, stateful runtime behavior.
Next-Generation Platform: Combines AutoGen's
simple abstractions with Semantic Kernel's enterprise features
such as type safety, telemetry, and human-in-the-loop support.
🤖Agents
LLM-driven units that reason, call tools and MCP servers,
and produce responses with autonomous planning.
🧭Workflows
Graph-based execution for type-safe routing, checkpointing,
and coordinated multi-agent or function pipelines.
🧱Core Runtime
Sessions, memory context providers, middleware, and MCP
clients provide safe, stateful, extensible foundations.
🚀Why It Matters
Direct successor to Semantic Kernel + AutoGen with stronger
enterprise controls and explicit orchestration patterns.
Theory 3
Core Agent Loop: Think → Act → Observe → Respond
Agents run a continuous execution loop. Each cycle processes
inputs, selects actions, captures results, and decides whether to
respond or continue reasoning.
1
🧠
Think
Analyze context, reason about next step
2
⚡
Act
Call tool, invoke API, or query MCP
3
👁️
Observe
Capture output, evaluate results
4
💬
Respond
Output answer or loop back to Think
Repeat until goal achieved
Theory 4
Exposing Tools for Actionable AI
Tool calling converts static text generation into deterministic
Python execution. Register functions using @tool and the
SDK auto-generates JSON schemas for the LLM.
👤
User Query
"Check server health"
→
🤖
OpsAgent
Sends chat history + tool schemas to LLM
→
🧠
LLM Decides
Returns tool_call JSON with arguments
→
🔧
@tool Function
Executes check_health("api-01")
→
✅
Grounded Answer
Verified, deterministic result returned to user
@tool(approval_mode="never_require")
Theory 5
What is MCP (Model Context Protocol)?
An open standard that decouples agent hosts from tool
integrations. Instead of custom bindings, MCP servers publish
resources, tools, and prompts over JSON-RPC.
OpsAgent is an AI-powered operations & engineering assistant. It
helps developers troubleshoot issues, retrieve documentation via
MCP, analyze systems, and automate workflows — all grounded in
real data.
Before the Microsoft Agent Framework, which (two) Microsoft frameworks were used to build AI agents?
Answer
Common Microsoft approaches
Semantic Kernel and AutoGen.
Question
What's the difference between a session provider and a context (user memory) provider?
Answer
Session vs Context
Session providers (e.g., AgentSession / InMemoryHistoryProvider) store ephemeral chat history for the current interaction. Context (user memory) providers persist facts, preferences and user data across sessions using a database (SQLite, Cosmos DB) and are used to inject long‑term knowledge into the agent's context.
