We already review Ed Donner's LLM engineering course on this site, and this is the second leg of what he now sells as a three-part AI engineer path: Core Track, Agentic Track and Production Track. The Agentic Track used to be called The Complete Agentic AI Engineering Course, and plenty of people still search for it under that name. When I checked the page it showed 422,868 learners, a 4.7 average from 51,365 ratings, 20 hours 52 minutes of video and a last update in September 2026. The structure is six weeks. Week 1 is foundations, then one week each on the OpenAI Agents SDK, CrewAI, LangGraph and AutoGen, and a final week on MCP.
There are eight projects along the way: a digital twin that answers questions about your career, a sales outreach agent, a deep research team, a stock picker, a four-agent engineering team running in Docker, a browser sidekick built on LangGraph, an agent that creates other agents, and a trading floor capstone. Week 1 is the reason I recommend the course. Donner explains what tool calling really is, then has you write the agent loop yourself with a plain while loop, handle multiple tool calls, and deploy the result. He covers the standard design patterns (chaining, routing, orchestrator and workers), compares model providers through one OpenAI-compatible client, runs local models with Ollama, and uses one model to judge the others. If you only did this week you would understand agents better than most people who have finished a framework tutorial. The framework weeks are where my enthusiasm drops a little.
I understand why he covers four. Students want to know which one to pick, and seeing the same ideas expressed four ways does teach you that the frameworks are thin layers over the same loop. But a week is not long enough to hit the problems that matter in any of them, and the AutoGen material in particular feels like it belongs to an earlier moment now that Microsoft has shifted its effort to a successor. If I were taking this today I would watch that week at double speed. The MCP week is strong and current.
You build servers and clients, connect several together and see what happens to an agent when it has dozens of tools available. The capstone, four trader agents with six MCP servers and 44 tools, is a good stress test of the idea, even if I would not let it near a brokerage account. What is missing is the unglamorous half of agent work. There is a lecture on risks, guardrails and evals in Week 1 and guardrails come back with the Agents SDK, but there is no sustained treatment of how you measure whether an agent is any good or what you do when it fails one run in twenty. In my experience that is most of the job.
Donner's Production Track is meant to pick this up, which is fair, but it means this course on its own produces people who can build impressive demos. He is an energetic teacher, sometimes relentlessly so. I find it easier to sit through 21 hours with someone who is clearly enjoying himself. The page says the course works for all backgrounds and there is a folder of self-study labs for beginners. I would not take that at face value. You need to be comfortable in Python, with a terminal and with git, or the setup alone will eat your first week. On cost, it is a Udemy course, so wait for the sale price or use a subscription.
Compared with Eden Marco's LangChain course this is broader and less deep. Compared with the free Hugging Face agents course it is better structured and more fun. For the money it is the agent course I would point a working developer to first.