Building AI Agents with C# .NET and Semantic Kernel training course

The Core Brain: Semantic Kernel & Function Logic

Focus: Building intelligent agents that use tools, plan steps, and reason iteratively.

1.1 Kernel Foundations

• Building a multi‑model Kernel using KernelBuilder
• Registering the Kernel via .NET Dependency Injection
• Structuring AI services for reasoning, summarization, and domain‑specific tasks

1.2 Native C# Plugins

• Writing C# classes as callable AI tools
• Using [KernelFunction] and [Description] to expose capabilities
• Passing strongly typed objects and records into LLM‑invoked functions

1.3 Stepwise Planning

• Implementing the FunctionCallingStepwisePlanner
• Understanding the reasoning loop: Goal → Plan → Tool Execution → Observation → Next Step
• Handling planner dead‑ends and misfires

1.4 Debugging Agent Failures

A new module focused on real‑world failure modes:

• Infinite loops
• Incorrect tool selection
• Planner hallucinations
• Bad arguments passed to functions
• Observing and correcting reasoning drift

Project — System Admin Agent

Build an agent that:

• Reads local logs
• Checks CPU usage via C# methods
• Diagnoses issues and suggests fixes
• Recovers gracefully from planner mistakes

The Digital Workforce: Multi‑Agent Systems

Focus: Designing specialized agents that collaborate, hand off tasks, and maintain memory.

2.1 Agent Personas & System Instructions

• Creating ChatCompletionAgent instances
• Defining backstories, constraints, and role boundaries
• Building personas such as:
  • Researcher
  • Writer
  • Critic

2.2 Orchestration Patterns

• Sequential “waterfall” flows (Researcher → Writer → Critic)
• Group chat dynamics
• Agent handoffs when a task exceeds one agent’s expertise
• Designing workflows that mimic real digital teams

2.3 Memory & State

• Using ChatHistory for short‑term memory
• Implementing long‑term memory with vector stores (Azure AI Search, Qdrant)
• Retrieval‑Augmented Planning (new module):
  • Agents retrieve relevant knowledge
  • Use it to shape their plan
  • Reduce hallucination and improve accuracy

2.4 Human‑in‑the‑Loop Patterns

• When agents should ask for clarification
• Approval workflows
• Escalation patterns for ambiguous or high‑risk tasks

Project — Content Pipeline

A multi‑agent workflow where:

• The Researcher gathers information
• The Writer produces a draft
• The Critic enforces brand voice and correctness
• The system can pause to ask the human for approval