Every few days this month, another "just dropped: a free course on building AI agents" post rockets around X. Three of them keep resurfacing — from Anthropic, Andrew Ng, and Google — and the reflex is to assume they're interchangeable and pick whichever has the most retweets. They aren't interchangeable. They teach genuinely different things, and picking wrong wastes the exact weekend you set aside to learn this.
The short answer: Take Agent Skills with Anthropic if you're building on Claude and want reusable workflows you can ship this week. Take Andrew Ng's Agentic AI if you want vendor-neutral fundamentals that survive a framework switch. Take Google's "agentic engineering from scratch" if you learn by wiring the loop and the protocol yourself. If you have time for two, do Ng's patterns first, then one hands-on course.
The three, side by side#
| Course | Who | Length | What you leave with | Best if you |
|---|---|---|---|---|
| Agent Skills with Anthropic | Anthropic (Elie Schoppik), on DeepLearning.AI | ~1 hr, hands-on | Custom + pre-built Agent Skills across Claude.ai, Claude Code, the API, and the Agent SDK; skills + MCP + subagents | Ship on Claude and want reusable, on-demand workflows now |
| Agentic AI | Andrew Ng, on DeepLearning.AI | Longer, structured | The agentic design patterns — reflection, tool use, planning, multi-agent — vendor-neutral | Want fundamentals that outlast any one framework |
| Agentic engineering from scratch | Google (free video walkthrough) | ~1 hr, watch-and-build | Your first agent, agent memory, long-running loops, your own MCP (MCP vs API), multi-agent | Learn by building the machinery yourself |
What each one is actually for#
Agent Skills with Anthropic is the narrowest and the most immediately useful. Taught by Anthropic's Elie Schoppik on DeepLearning.AI, it teaches one concrete mechanism: Agent Skills — folders of instructions that give an agent on-demand knowledge and workflows, so you stop re-explaining the same process every session. It walks through Anthropic's pre-built skills (Excel, PowerPoint, skill creation), building your own following best practices, and running them across Claude.ai, Claude Code, the Claude API, and the Agent SDK — then combining skills with MCP and subagents. If you're already on Claude, this is the course you can turn into shipped work fastest. (We went deeper on the founder angle in what founders should take from Anthropic's Agent Skills course.)
Andrew Ng's Agentic AI is the foundations track. It's built around the durable agentic design patterns — reflection (the agent critiques its own output), tool use, planning, and multi-agent collaboration — that Ng has taught for a while and that hold regardless of whether you end up on LangGraph, CrewAI, the Claude Agent SDK, or your own loop. It's the least likely of the three to go stale, because it's about how agents are shaped, not which SDK is hot this quarter. If you only internalize one thing before writing code, make it these patterns.
Google's "agentic engineering from scratch" is the build-it-yourself path — a free, roughly one-hour walkthrough that keeps resurfacing on X. It starts from nothing and wires up: your first agent, agent memory (short-term, persistent, long-term), agentic loops for long-running agents, building your own MCP and when MCP beats a plain API, and multi-agent systems. It's the most hands-on-with-the-machinery of the three. That makes it a great second course — once you know the patterns, watching the loop and the protocol get built by hand is what makes them stick.
The honest routing#
- Shipping on Claude this week? Anthropic's skills course. Shortest path from watching to a reusable skill in your repo.
- Want knowledge that survives your next framework migration? Ng's patterns. Frameworks churn; reflection-tool-use-planning-multi-agent doesn't.
- Learn by building? Google's from-scratch walkthrough, ideally after Ng's.
- Have a weekend? Ng first (the map), then one hands-on course (the terrain). Skip trying to do all three back-to-back — the hands-on courses overlap enough that a second one has diminishing returns.
The pattern worth noticing across all three: every one of them now treats skills, memory, MCP, and multi-agent as the core curriculum. Two years ago an "AI course" meant prompting. The free courses builders are sharing today assume you're building a system — which is a decent signal of where the floor for "knows how to build an agent" has moved. Whichever you pick, that's the level you're being asked to reach.
