Forward deployed engineer was a Palantir job title for most of its life. In 2026 the AI labs and a lot of startups adopted it for engineers who sit with customers and make agents work inside their systems, and searches for a course on it went from a few dozen a month to several hundred over the summer. Ed Donner got there quickly. The page is tagged Hot and New, with 16,198 learners, a 4.8 average from 1,169 ratings, 21 hours 25 minutes of video across 124 lectures, and an update in October 2026. The three weeks are labelled become an AI engineer, become an AI architect and become an AI FDE.
Ignore the labels. What matters is that the whole course is built on one project: a claims agent for a made-up insurer called UnderCover. You start with a dataset of claims and the decisions human handlers made, and everything you build is measured against it. Week 1 moves fast. There is an agent loop written from scratch that gets a model to solve a maze, a comparison of frontier models on the claims data, and a day where Claude Code plans and builds the pilot in phases with success criteria.
That day includes hardening the repository with linters, type checks, tests and a pre-commit hook before the agent is let loose, which is the right order and one I rarely see taught. Day 4 is a detour into running models locally: VRAM, quantization, file formats, Ollama against hosted providers, and what each option costs per claim. Day 5 compares building your own agent harness with using the Claude Agent SDK or LangChain's Deep Agents. Week 2 is where this course separates from his earlier ones. From the syllabus and the outcomes list it covers evals and tracing in Langfuse, agentic RAG over MCP for long-term memory, and LangGraph workflows with durable state, crash recovery and a human approval step.
In my review of the Agentic Track I complained that evals got a lecture when they needed a week. Here they are used to make architecture decisions, which is how it works on a real engagement. You change something, you rerun the claims, and the numbers tell you whether it helped. Week 3 covers swarms and using one model to judge another, then the client-facing material: discovery, metrics, governance, responsible AI and handover. It also has the security content, including prompt injection and the so-called lethal trifecta of private data, untrusted input and the ability to send data out.
The final build deploys the LangGraph agents to AWS Bedrock AgentCore using its runtime, gateway, policy and memory services. I have two criticisms. The first is the framing. The description opens with postings up 1,000 percent, salaries reaching $400,000 and 30,000 openings. Donner says he read more than 300 job descriptions to pick the topics, and I believe that, because the syllabus matches what those adverts ask for.
But the difficult part of a forward deployed job is not LangGraph. It is working out what a client's operations team does all day, finding the one process worth automating, and getting security and legal to agree to it. Udemy has added a role play feature and the course uses it once. That is a gesture in the right direction and no substitute for doing it. The second is AgentCore.
Choosing a concrete platform is better than waving at deployment in general, and AWS is a reasonable bet. It is still a young product from one vendor, and if your employer or client is on Azure or Google Cloud that part will not transfer directly. The requirements line says some experience with Python and LLMs is ideal but not required. I would say required. Week 1 has enough setup help for a beginner to survive, but durable state and gateway policies are not beginner topics. He says every paid API has a free alternative and you need not spend anything, and from the amount of OpenRouter and Ollama material I think that is true.
If you already own the Agentic Track you will see familiar ground in week 1. For everyone else who can already code, I think this is now the best starting point in his catalogue, because it teaches you to measure an agent before it teaches you to multiply them.