Building with the Claude API: Anthropic's free course covers half the engineering of the FDE job

The Forward Deployed Engineer posting that Anthropic’s Applied AI team published on 28 April 2026 asks candidates to deliver very specific things to clients: MCP servers, sub-agents, agent skills. Anthropic’s own free course has dedicated modules on MCP and on agent workflows. The syllabus and the job description overlap in quite a few places.

The same posting, however, expects FDEs to spend 25 to 50% of their time at the client’s site, building the product alongside them. No video can teach that part.

If you are a developer weighing a move into FDE work, the course is worth your time. You just need to know which half of the job it covers and which half you will have to learn on your own.

Who is the course for?

“Building with the Claude API” is made by Anthropic, hosted on Skilljar as part of Anthropic Academy, and free to enrol in. Anthropic Academy describes itself as the place where Anthropic’s education team publishes free courses and guides. The course page describes it as a video course that teaches developers to integrate Claude into applications through the Anthropic API.

The entry bar is low: basic Python and JSON, with no prior LLM experience needed. Notably, the target audience explicitly includes software engineers moving into AI application development. That is almost exactly the profile of someone trying to break into FDE.

The course page does not state a duration, so set your own schedule, for example one module a week plus a small exercise to go with it.

Which modules should you take first?

Start with the Prompt evaluation module, before RAG and before agents. It teaches you to build a test dataset and then grade the results with both a model and code. At a client’s site, this is the tool that lets you prove the system works with numbers rather than gut feeling.

Picture a client that is an insurance company wanting Claude to extract information from claims files. You write 20 sample files. Code checks whether the output is valid JSON and whether all required fields are present.

The model grades what code cannot, such as whether the summary has invented details. If 17 of the 20 files pass both rounds, you have a number to bring into the meeting with the client, and the remaining 3 tell you what to fix next.

After evals, take Tool use. The module runs from schemas to multi-turn conversations with multiple tools, and adds the text edit tool and the web search tool. If time is short, you can skim those two built-in tools.

Spend your time instead on writing schemas and handling multi-turn conversations with multiple tools, because every tool-based system you build for a client will need both skills.

Next comes RAG and Agentic Search: chunking, embeddings, BM25 and multi-index pipelines. Almost every client has a pile of internal documents and wants to ask questions of it. Combining BM25 with embeddings lets you catch both exactly typed product codes and questions phrased in many different ways.

Three ideas worth keeping

The first is that evals need two layers of grading. Code grades whatever can be measured mechanically; a model grades the rest. Drop either layer and you are fooling yourself.

The second comes from the Agents and workflows module, with lessons on parallelization, chaining and routing, and a separate lesson comparing workflows with agents. Choose the simplest structure that still solves the problem. A document-review process with fixed steps needs nothing more than chaining. Only when the steps cannot be predicted do you need an agent that decides for itself.

The third is MCP. The module covers clients, servers, tools, resources and prompts, and an MCP server is one of the things Anthropic’s posting says FDEs must deliver to clients. So learn MCP as a way of packaging a deliverable, not as a protocol you study just to say you know it.

The half the course does not teach

Gergely Orosz of The Pragmatic Engineer defines an FDE as an engineer who rotates between sitting with client teams and working on the core product team. The role originated at Palantir in the early 2010s. Google’s FDE posting describes someone who builds directly inside the client’s team, connecting cutting-edge AI products with production reality.

Anthropic’s posting likewise requires working inside client systems, as well as identifying and codifying recurring deployment patterns. No module teaches you how to read an unfamiliar codebase in two days, or how to turn your third deployment into a template the product team can reuse.

You can fill this gap with a personal project. Pick a hypothetical problem, say the customer service department of a retail chain. Build an MCP server that wraps their mock API, write an eval suite, then write a one-page description of the pattern that the next deployment could reuse.

Put exactly that on your CV: the deliverable, the eval numbers and the pattern you extracted. When reading a job description, look for phrases such as “customer systems” or “deployment patterns” to see which half of you the company needs.

The course helps you write the code the client needs. Whether the client invites you back for more depends on the other half of the job.