Image by Author | Canva & ChatGPT

# Introduction

 
GitHub has become the go-to platform for beginners eager to learn new programming languages, concepts, and skills. With the growing interest in agentic AI, the platform is increasingly showcasing real projects that focus on “agentic workflows,” making it an ideal environment to learn and build.

One notable resource is microsoft/ai-agents-for-beginners, which features a 12-lesson course covering the fundamentals of building AI agents. Each lesson is designed to stand on its own, allowing you to start at any point that suits your needs. This repository also offers multi-language support, ensuring broader accessibility for learners. Each lesson in this course includes code examples, which can be found in the code_samples folder.

Moreover, this course uses Azure AI Foundry and GitHub Model Catalogs for interacting with language models. It also incorporates several AI agent frameworks and services like Azure AI Agent Service, Semantic Kernel, and AutoGen.

To facilitate your decision-making process and provide a clear overview of what you will learn, we will review each lesson in detail. This guide serves as a helpful resource for beginners who might feel uncertain about choosing a starting point.

# 1. Intro to AI Agents and Agent Use Cases

 
This lesson introduces AI agents — systems powered by large language models (LLMs) that sense their environment, reason over tools and knowledge, and act — and surveys key agent types (simple/model-based reflex, goal/utility-based, learning, hierarchical, and multi-agent systems (MAS)) through travel-booking examples.

You will learn when to apply agents to open-ended, multi-step, and improvable tasks, and the foundational building blocks of agentic solutions: defining tools, actions, and behaviors.

# 2. Exploring AI Agentic Frameworks

 
This lesson explores AI agent frameworks with pre-built components and abstractions that let you prototype, iterate, and deploy agents faster by standardizing common challenges and boosting scalability and developer efficiency.

You will compare Microsoft AutoGen, Semantic Kernel, and the managed Azure AI Agent Service, and learn when to integrate with your existing Azure ecosystem versus using standalone tools.

# 3. Understanding AI Agentic Design Patterns

 
This lesson introduces AI agentic design principles, a human-centric user experience (UX) approach for building customer-focused agent experiences amid the inherent ambiguity of generative AI.

You will learn what the principles are, practical guidelines for applying them, and examples of their use, with an emphasis on agents that broaden and scale human capacities, fill knowledge gaps, facilitate collaboration, and help people become better versions of themselves through supportive, goal-aligned interactions.

# 4. Tool Use Design Pattern

 
This lesson introduces the tool-use design pattern, which allows LLM-powered agents to have controlled access to external tools such as functions and APIs, enabling them to take actions beyond just generating text.

You will learn about key use cases, including dynamic data retrieval, code execution, workflow automation, customer support integrations, and content generation/editing. Additionally, the lesson will cover the essential building blocks of this design pattern, such as well-defined tool schemas, routing and selection logic, execution sandboxing, memory and observations, and error handling (including timeout and retry mechanisms).

# 5. Agentic RAG

 
This lesson explains agentic retrieval-augmented generation (RAG), a multi-step retrieval-and-reasoning approach driven by large language models (LLMs). In this approach, the model plans actions, alternates between tool/function calls and structured outputs, evaluates results, refines queries, and repeats the process until achieving a satisfactory answer. It often uses a maker-checker loop to enhance correctness and recover from malformed queries.

You will learn about the situations where agentic RAG excels, particularly in correctness-first scenarios and extended tool-integrated workflows, such as API calls. Additionally, you will discover how taking ownership of the reasoning process and using iterative loops can enhance reliability and outcomes.

# 6. Building Trustworthy AI Agents

 
This lesson teaches you how to build trustworthy AI agents by designing a robust system message framework (meta prompts, basic prompts, and iterative refinement), enforcing security and privacy best practices, and delivering a quality user experience.

You will learn to identify and mitigate risks, such as prompt/goal injection, unauthorized system access, service overloading, knowledge-base poisoning, and cascading errors.

# 7. Planning Design Pattern

 
This lesson focuses on planning design for AI agents. Start by defining a clear overall goal and establishing success criteria. Then, break down complex tasks into ordered and manageable subtasks.

Use structured output formats to ensure reliable, machine-readable responses, and implement event-driven orchestration to address dynamic tasks and unexpected inputs. Equip agents with the appropriate tools and guidelines for when and how to use them.

Continuously evaluate the outcomes of the subtasks, measure performance, and iterate to improve the final results.

# 8. Multi-Agent Design Pattern

 
This lesson explains the multi-agent design pattern, which involves coordinating multiple specialized agents to collaborate toward a shared goal. This approach is particularly effective for complex, cross-domain, or parallelizable tasks that benefit from the division of labor and coordinated handoffs.

In this lesson, you will learn about the core building blocks of this design pattern: an orchestrator/controller, role-defined agents, shared memory/state, communication protocols, and routing/hand-off strategies, including sequential, concurrent, and group chat patterns.

# 9. Metacognition Design Pattern

 
This lesson introduces metacognition, which can be understood as “thinking about thinking,” for AI agents. Metacognition allows these agents to monitor their own reasoning processes, explain their decisions, and adapt based on feedback and past experiences.

You will learn planning and evaluation techniques, such as reflection, critique, and maker-checker patterns. These methods promote self-correction, help identify errors, and prevent endless reasoning loops. Additionally, these techniques will enhance transparency, improve the quality of reasoning, and support better adaptation and perception.

# 10. AI Agents in Production

 
This lesson demonstrates how to transform “black box” agents into “glass box” systems by implementing robust observability and evaluation techniques. You will model runs as traces (representing end-to-end tasks) and spans (petitions for specific steps involving language models or tools) using platforms like Langfuse and Azure AI Foundry. This approach will enable you to perform debugging and root-cause analysis, manage latency and costs, and conduct trust, safety, and compliance audits.

You will learn what aspects to evaluate, such as output quality, safety, tool-call success, latency, and costs, and apply strategies to enhance performance and effectiveness.

# 11. Using Agentic Protocols

 
This lesson introduces agentic protocols that standardize the ways AI agents connect and collaborate. We will explore three key protocols:

Model Context Protocol (MCP), which provides consistent, client-server access to tools, resources, and prompts, functioning as a “universal adapter” for context and capabilities.

Agent-to-Agent Protocol (A2A), which ensures secure, interoperable communication and task delegation between agents, complementing the MCP.

Natural Language Web Protocol (NLWeb), which enables natural-language interfaces for websites, allowing agents to discover and interact with web content.

In this lesson, you will learn about the purpose and benefits of each protocol, how they enable large language models (LLMs) to communicate with tools and other agents, and where each fits into larger architectures.

# 12. Context Engineering for AI Agents

 
This lesson introduces context engineering, which is the disciplined practice of providing agents with the right information, in the right format, and at the right time. This approach enables them to plan their next steps effectively, moving beyond one-time prompt writing.

You will learn how context engineering differs from prompt engineering, as it involves ongoing, dynamic curation rather than static instructions. Additionally, you will understand why strategies such as writing, selecting, compressing, and isolating information are essential for reliability, especially given the limitations of constrained context windows.

# Final Thoughts

 
This GitHub course provides everything you need to start building AI agents. It includes comprehensive lessons, short videos, and runnable Python code. You can explore topics in any order and run samples using GitHub Models (available for free) or Azure AI Foundry.

Additionally, you will have the opportunity to work with Microsoft’s Azure AI Agent Service, Semantic Kernel, and AutoGen. This course is community-driven and open source; contributions are welcome, issues are encouraged, and it is licensed for you to fork and extend.
 
 

Abid Ali Awan (@1abidaliawan) is a certified data scientist professional who loves building machine learning models. Currently, he is focusing on content creation and writing technical blogs on machine learning and data science technologies. Abid holds a Master’s degree in technology management and a bachelor’s degree in telecommunication engineering. His vision is to build an AI product using a graph neural network for students struggling with mental illness.