What are AI Agents

Here’s your complete study guide for all 10 lectures, with detailed notes and free resources for each.

Lecture 1 – What are AI Agents—

📌 Lecture 1 — What are AI Agents

Core concepts: An AI agent is a system where an LLM acts as a reasoning engine that can perceive its environment, make decisions, and take actions (use tools, call APIs, run code) in a loop to complete a goal — rather than just responding once to a prompt.

Key ideas to understand:

• Perception → Reasoning → Action loop (the agent cycle)
• Agents vs. simple LLM calls: agents have memory, tools, and goals
• Types: single agents, multi-agent systems, autonomous vs. human-in-the-loop
• Tools/function calling as the bridge between language and action

Free resources:

• Anthropic’s “What are agents?” overview — very practical, explains agents clearly with patterns
• Hugging Face AI Agents Course – Unit 1 — free, beginner-friendly, interactive
• LangChain conceptual docs on agents — solid conceptual grounding

📌 Lecture 2 — Inside the Brain of AI Agents: How LLMs Work

Core concepts: Transformers, tokenization, attention mechanism, how LLMs predict the next token, context windows, and why this matters for agent reasoning.

Key ideas:

• Text → tokens → embeddings → transformer layers → output probabilities
• Self-attention: how the model “looks” at all previous tokens simultaneously
• Temperature and sampling: how randomness affects generation
• Context window limits and why they constrain agents
• In-context learning vs. fine-tuning

Free resources:

• 3Blue1Brown: “But what is a neural network?” (YouTube) — best visual intuition for neural nets
• Andrej Karpathy: “Let’s build GPT from scratch” (YouTube) — the gold standard deep dive, ~2 hours
• Jay Alammar: “The Illustrated Transformer” — outstanding visual walkthrough of attention
• Hugging Face NLP Course – Chapter 1 — free and hands-on

📌 Lecture 3 — How Agents Really Work: The ReAct Framework

Core concepts: ReAct (Reasoning + Acting) is the dominant pattern for agents. The model interleaves Thought → Action → Observation steps in a loop, reasoning about what to do next based on tool results.

Key ideas:

• The Thought/Action/Observation loop
• How tool calls are embedded in model output
• Why chain-of-thought reasoning improves tool use
• ReAct vs. simple tool-calling vs. chain-of-thought

Free resources:

• Original ReAct paper (arXiv) — free, readable, ~10 pages
• Lilian Weng: “LLM Powered Autonomous Agents” (blog) — the best single blog post on agent architectures
• Hugging Face Agents Course – Unit 2: ReAct — free, code-along

📌 Lecture 4 — Overview of AI Agentic Frameworks

Core concepts: The landscape of tools for building agents — from raw code (calling APIs yourself), to frameworks (LangChain, LlamaIndex, CrewAI), to low-code (n8n, Flowise), to no-code (Zapier AI, Make).

Key ideas:

• Trade-offs: control vs. speed of development
• When to use a framework vs. roll your own
• Categories: orchestration frameworks, RAG frameworks, workflow automation tools
• Cost and latency implications

Free resources:

• a16z: “The Emerging LLM App Stack” (blog) — great map of the ecosystem
• Flowise docs — free low-code agent builder, open source
• Zapier AI overview (YouTube) — search for beginner tutorials

📌 Lecture 5 — Smolagents: The Simplest Agent Coding Library

Core concepts: Smolagents (by Hugging Face) is a minimal Python library for building agents. Its design philosophy is simplicity — agents are just Python code with tool-decorated functions.

Key ideas:

• CodeAgent vs. ToolCallingAgent
• Defining tools with @tool decorator
• The model-agnostic design (works with any LLM)
• How it differs from heavier frameworks like LangChain

Free resources:

• Smolagents official docs — free, comprehensive
• Hugging Face Agents Course – Unit 3: Smolagents — free, code-along with notebooks
• Smolagents GitHub — examples folder is very instructive

📌 Lecture 6 — Building Multi-Agent Frameworks and Browser Agents

Core concepts: Multi-agent systems have a manager/orchestrator agent that delegates tasks to specialized sub-agents. Browser agents can control a real web browser to perform web tasks.

Key ideas:

• Orchestrator → worker agent patterns
• How agents communicate (message passing, shared memory)
• Browser automation: Playwright/Selenium under the hood
• Challenges: error recovery, hallucinated actions, infinite loops

Free resources:

• Smolagents multi-agent guide — free official tutorial
• Browser Use library (GitHub) — open-source browser agent library with examples
• Hugging Face Agents Course – Unit 4: Multi-agents
• Microsoft Playwright docs — underlying browser control tech, free

📌 Lecture 7 — Agentic RAG Using LlamaIndex

Core concepts: RAG (Retrieval-Augmented Generation) grounds LLMs with external knowledge. Agentic RAG goes further — the agent decides when and how to retrieve, can query multiple sources, re-rank, and reason over retrieved chunks.

Key ideas:

• Basic RAG pipeline: chunk → embed → store → retrieve → generate
• Query engines, retrievers, and routers in LlamaIndex
• Agent-as-query-engine: tools for different data sources
• Sub-question decomposition for complex queries

Free resources:

• LlamaIndex docs: “Building Agentic RAG” — free, official
• LlamaIndex RAG bootcamp (YouTube) — free video walkthrough
• Jerry Liu (LlamaIndex founder) talks on YouTube — very concept-heavy, free
• LlamaIndex GitHub notebooks — hands-on code

📌 Lecture 8 — LangGraph in 100 Minutes

Core concepts: LangGraph models agent workflows as stateful graphs — nodes are functions/LLM calls, edges are transitions. This makes complex, cyclical, multi-step agents much easier to control and debug.

Key ideas:

• StateGraph: defining state, nodes, and edges
• Conditional edges for branching (if agent needs tool → call tool, else → end)
• Checkpointer for persistence and memory across turns
• Human-in-the-loop with interrupt_before
• How it differs from LangChain’s older AgentExecutor

Free resources:

• LangGraph official docs — free, excellent tutorials
• LangGraph “Get started” tutorial — step-by-step, free
• LangChain YouTube channel — tons of free LangGraph walkthroughs
• LangGraph Academy (free tier) — structured free course

📌 Lecture 9 — Building Agents Using CrewAI

Core concepts: CrewAI is a framework for orchestrating teams of role-playing AI agents. You define agents with roles, goals, and backstories, then assign tasks — agents collaborate, delegate, and produce outputs.

Key ideas:

• Agent (role + goal + backstory), Task, Crew objects
• Sequential vs. hierarchical process modes
• Tools in CrewAI (built-in web search, custom tools)
• Memory types: short-term, long-term, entity memory

Free resources:

• CrewAI official docs — free, comprehensive
• CrewAI crash course (YouTube – Brandon Hancock) — very popular free tutorial
• CrewAI GitHub examples — many real-world project templates
• DeepLearning.AI: “Multi AI Agent Systems” (free) — official short course, free

📌 Lecture 10 — n8n and Agentic Automations

Core concepts: n8n is an open-source workflow automation tool (like Zapier, but self-hostable) with native AI agent nodes. You can build LLM-powered automations with a visual canvas — no code required.

Key ideas:

• Triggers, nodes, and connections in n8n
• The AI Agent node: connecting to OpenAI/Anthropic
• Memory nodes (window buffer, Postgres, Redis)
• Tool nodes: HTTP request, code, Google Sheets, email
• Hosting: n8n Cloud (free tier) vs. self-hosted

Free resources:

• n8n official docs — free, well-organized
• n8n AI & LangChain docs — specific to agent use cases
• n8n YouTube channel — dozens of free tutorial videos
• n8n free cloud tier — sign up and build right in the browser
• Cole Medin’s n8n tutorials (YouTube) — excellent in-depth free walkthroughs

🗺️ Suggested Learning Order

Follow the curriculum order — it’s well-sequenced. Lectures 1–3 are pure theory and take 1–2 days each. Lectures 4–6 get hands-on with code. Lectures 7–9 are each a separate framework requiring practice projects. Lecture 10 is no-code and a good “Friday afternoon” project.

For each framework lecture (5–10), the best approach is: read the docs → watch one tutorial → build one small project of your own. That cycle will reinforce the concepts far better than passive reading alone.