Ed Donner’s Free LLM Engineering and Agentic AI Courses

Yes, with a clear boundary. The repositories contain the full code curriculum, setup guides, challenges, and the resources page, which is everything a self-directed learner needs to work through the material. The videos add the narration, the demonstrations of intent behind design decisions, and the pacing of a taught course, and the author’s framing makes clear the repos and videos are companions with the repos as the artifact. Learners who work well from code plus guides will lose little. Learners who need a voice explaining why the next cell matters will feel the gap in weeks where the concepts stack quickly, particularly the transition into agentic orchestration. A workable hybrid: run the repos free, and if a specific week stalls, that is the week to buy or watch selectively.

Because the first hour decides whether a learner continues, and a billing setup flow kills momentum faster than any conceptual difficulty. The Ollama-first opening guarantees a working result before any account creation, payment method, or API key exists, which removes the three most common Day 1 abandonment points at once. It also plants a mental model that pays off later: models are software you can run, not services you must rent, and the course returns to that framing whenever costs come up. The technical caveat the README flags carefully is model size, steering learners to llama3.2 rather than the much larger llama3.3, because a mismatched local model turns the instant-gratification play into a downloading-failure horror story, and the course catches that before the learner meets it. That level of operational care in a README is a reliable signal for how the rest of the course is maintained.

The refresh rebuilt the content around the current framework landscape rather than patching it, and the upgrade documentation addresses the three questions returning students actually have: what changed, why it needed changing, and how to migrate code from the middle of the old version. In a field where courses routinely break silently on library updates, publishing an explicit migration path is a maintenance standard worth recognizing. The substance of the refresh tracks the ecosystem’s consolidation: the OpenAI Agents SDK as a serious production path, MCP as the standard connective layer, and Google ADK joining the established CrewAI and LangGraph. New learners should simply start on the refreshed version; mid-course learners should read the migration notes before updating, since framework rewrites mid-project are exactly when half-migrated states bite.

The course’s breadth is a map, not a mandate, and depth in one framework beats surface contact with six. The choosing logic: if your employer or target employers standardize on one, that one, obviously. Otherwise, the OpenAI Agents SDK is the safe production default with first-party momentum, LangGraph rewards teams building stateful multi-step workflows and wanting explicit control over the graph, CrewAI remains the fastest path to multi-agent team metaphors, Pydantic AI suits Python shops that live on type safety, and Google ADK is the pick inside the Google ecosystem. MCP is not optional regardless of framework, since it is becoming how tools connect across all of them. A reasonable plan: work every week, but build your course project in the framework closest to your next twelve months, and treat the others as reading comprehension.

The challenges are woven into the weeks as stretch exercises after the taught material, sized for an evening rather than a weekend, and the invitation to submit solutions as pull requests is genuine: community contributions land in a dedicated folder with recognition. For learners, this is a low-stakes first open-source contribution, which many engineers otherwise postpone for years, and reviewing classmates’ alternative solutions to the same challenge is where much of the real learning hides, since the differences between correct implementations are exactly the design decisions that matter. For the course, the culture keeps the code honest across platforms, because a solution that only works on one operating system gets caught by the next learner on another. If you take one habit from the repos, take this one: finish a challenge, open the PR, and read two classmates’ solutions before merging your mental model shut, because the differences are the lesson.

A team should treat the split as a curriculum boundary. The foundation course is the onboarding artifact: uniform vocabulary, uniform tooling, and a cost profile that means a whole team can take it without an expense conversation, with the week 8 autonomous project doubling as an internal demo of what the team can now build. The agents course is the specialization artifact, and teams should not send everyone through all six frameworks. Pick the framework the team actually uses, have everyone do that week plus the MCP material deeply, and treat the remaining weeks as optional literacy, one person per framework who can advise when a use case suggests a different tool. Solo learners have the simpler problem: take both in order, but allow yourself to stop the agents course after your chosen framework plus MCP, because depth in one framework plus protocol fluency is the marketable combination, and the remaining weeks keep well as reference when a future role or project calls for them.

The guides are the part of the repos that first-time visitors under-read. Beyond the setup documents, the guides directory carries standalone notebooks on the recurring side-topics a course cannot interrupt itself for: working with the APIs, the free-alternative routes, version-control workflow for submitting solutions, and the introductory material that sets conventions. They are written to be usable without the course, which makes them a free reference library even for people who never start week 1. The resources page complements them with slides and supplemental videos, and the FAQ system answers the questions that actually recur, including the pragmatic ones, can this be taken without a programming background, what jobs it maps to, whether Gemini or free models substitute for OpenAI, with straight answers rather than marketing deflections. Together they form the maintenance layer of the course, the part that quietly decides whether a free curriculum still works in six months, and both are visibly tended, which is the strongest predictor that the courses will still be current when you finish them That maintenance signal also de-risks the sunk-cost worry with free courses generally: an actively tended repository rewards the effort you invest in learning its structure, because the structure is still there, still correct, when you return in six months for a refresher..

Ed Donner’s path into this material shapes how it reads. He built and sold tech companies before turning to AI education, and the courses carry that operator’s instinct: exercises end in artifacts, costs are stated before they surprise, and the syllabus optimizes for what a working engineer can absorb around a job rather than for academic completeness. The voice is direct and unusually present, with support channels answered personally, which readers of the repositories notice quickly. The combination lands as courses that respect both your time and your intelligence, a rarer pairing in free education than it should be.