A systematic, self-contained guide to LLM-powered AI agents, ordered the way the field should be learned: theory first, then core capabilities one at a time, then a from-scratch build, then production concerns.
You will learn what separates a workflow from an agent (and when each wins), the ReAct reasoning-and-acting loop every modern agent descends from, how tool calling actually works on the wire, planning and self-correction (Tree of Thoughts, Reflexion), short- and long-term memory including why naive RAG often fails, multi-agent orchestration with honest cost math, and production evaluation, observability, and guardrails.
The centerpiece is a weekend project: a complete minimal agent in pure Python, no frameworks, followed by a failure-mode lab where you break it on purpose: infinite loops, hallucinated tool calls, context bloat, and error compounding. Every code listing runs offline and is under a hundred lines.
Distilled from the field's primary sources, including the research canon (ReAct, Tree of Thoughts, Reflexion, Generative Agents) and production engineering accounts, and closed by a history of the field, a glossary, and a full bibliography. No survey filler. No framework dependency. Just the mechanics every agent is built from.
