Course introduction and an overview of building production-grade AI agents.

Examples: hello-world

The architecture and evolution of modern LLMs, from Markov chains to Transformers.

Examples: markov-chains-basic, markov-chains-books, tiny-llama-completions, tiny-llama-chat

From terminal to web: the client-server architecture behind an LLM chatbot.

Examples: markov-chains-temperature, ollama-completions, qwen-web-chat

Role-play, InstructGPT, and the principles of a good one-shot prompt.

Examples: few-shot, five-questions, five-questions-web

From chain-of-thought prompting to reasoning models. How In-Context Learning becomes Reasoning.

Agent harnesses and tool calling for LLM-based agents.

Examples: qwen-chat, qwen-tool-calling, gemini-hello-world, gemini-web-tool-calling

Guardrails on the agent loop and the OpenAI Agents SDK.

Examples: guardrails-password-challenge, guardrails-password-challenge-web, agents-sdk-chat, agents-sdk-web

Google ADK, Claude Agent SDK, LangGraph, and Pydantic AI.

Examples: adk-web

Deterministic checks, golden references, rubrics, gold and silver datasets, pass@k, and offline vs. online evals.

Examples: nyc-311-agent

Classic (semantic) RAG, retrieval as a tool call, reranking, hybrid search, and filesystem RAG.

Examples: magic-item-shop-classic-rag, magic-item-shop-semantic-rag, magic-item-shop-filesystem-rag

Agents that perform retrieval from SQL databases.

Examples: subway-survey-text2sql, subway-survey-text2sql-improved, nyc-311-agent, nyc-311-agent-improved

Tracing agents with OpenTelemetry and Langfuse.

Examples: nyc-311-agent-langfuse

Play five questions against Gemini while watching its context, then rewind and ask again to see that no object was ever chosen.

A web chatbot with a weather tool that shows every tool call the harness runs before the answer.

Get the model to reveal a password across seven levels, each stacking a new guardrail on the last.