A Google Developer Expert just published the most detailed guide to becoming an AI engineer I have seen. Not a pile of lessons, but a separate route for each specialty, and the course walks you through it. 523 lessons across 20 phases, free and MIT licensed. Linear algebra at the bottom, multi-agent swarms at the top, around 342 hours end to end. Code in Python, TypeScript, Rust and Julia. You never start at lesson one. You pick a destination and it hands you the path: > New to all of it, phase 0. > Already know Python, phase 1. > Want to ship production LLM apps, phase 11. > Want agents, phase 14. > Want MCP or Agent Skills, phase 13. There is a route for using coding agents on real repositories, one for deciding what to build before you build it, and one for Claude certification prep. Every lesson runs the same shape. You write the algorithm from raw math with no frameworks first, then run the same thing through PyTorch. By the time the library appears you already know what it does. And each lesson leaves something behind: a prompt, a skill, an agent or an MCP server. One command installs them into Claude, Cursor, Codex or anything reading SKILL.md.