FDE Masterclass

The role that broke out of Palantir and now defines hiring at Anthropic, OpenAI and the labs around them — and the three jobs it is most often confused with.

1. 1What an FDE Actually Does9 minThe role that broke out of Palantir and is now defining hiring at Anthropic, OpenAI and the labs around them — and what makes it different from every job title it is often confused with.
2. 2The Companies Hiring FDEs9 minWho is actually hiring Forward Deployed Engineers in 2026, what each company emphasises in its job description, and how to read a JD so you know which part of the craft a given role will press hardest on.
3. 3FDE vs SWE, SE and AI Engineer8 minThe three job titles an FDE role is most often confused with on a careers page, and the one test that separates them — who owns what, after whom, and with what access to the customer's codebase.

Tokens, context, cost and sampling — the mechanics that every later chapter assumes, taught with the honesty that an FDE needs on a customer call.

1. 4Tokens, Context and Cost9 minThe three numbers that drive every conversation an FDE has with a customer about an LLM — what a token actually is, what a context window actually holds, and how the pricing on a vendor page becomes a per-customer unit economics story.
2. 5The Model Families and When Each Is Right9 minFrontier, mid-tier, small and reasoning models as four separate tools — and the single rule that stops an FDE from reaching for the biggest one on reflex.
3. 6Sampling, Temperature and the Honest Limits of Determinism8 minTemperature, top-p and seed as the three sampling knobs — what each one does, what determinism you can buy with them, and the harder truth about what you cannot.
4. 7What Evaluation Even Means8 minEvaluation as a measurement problem — the three questions an eval answers, why cheap evals beat clever ones, and why the rest of this course treats the subject as its spine.

Scoping the prompt, binding the output to a schema, and prompt caching as the single cheapest lever an FDE has on the cost conversation.

1. 8Scoping a Prompt9 minWhat belongs in the prompt, what belongs in retrieval, and what belongs in a tool — the single division most FDEs get wrong on their first customer build.
2. 9Structured Output and Tool Schemas10 minBinding a model's output to a JSON schema, and the quiet surprise that a tool call is the same mechanism with a side effect. The failure modes a demo hides and production surfaces.
3. 10Prompt Caching as a Cost Lever8 minThe cheapest lever an FDE has in the customer cost conversation, how it actually works, and the cache-key mistake that turns a cost win into a data leak.

What retrieval is actually buying you, how to chunk without lying to yourself, hybrid search and reranking, and the failure modes a demo never shows.

1. 11What RAG Actually Is8 minRetrieval-augmented generation as a pipeline, not a product — what retrieval buys you over a longer context, and the first discipline of knowing when you do not need it at all.
2. 12Chunking Honestly9 minHow a corpus gets sliced into passages, why the quiet failure mode of RAG lives here, and what a bad chunk boundary does to a judge rubric you thought was calibrated.
3. 13Hybrid Search and Reranking9 minDense and sparse retrieval as complements, score fusion as the glue, and rerankers as the step that turns a plausible top-5 into one the model can actually answer from.
4. 14Grounding and Citation8 minBinding the generated answer to the retrieved passages, measuring groundedness as its own property, and the audit trail that turns a chatbot into something a compliance team will sign.

The agentic loop as an FDE uses it, authoring an MCP server a customer can trust, sub-agents and skills, and the coordinator-plus-specialists shape of production agent teams.

1. 15The Agentic Loop for FDEs8 minGather, act, verify — the loop every modern FDE is deploying — but with the customer-site angle a tutorial never has, and the stopping conditions that keep an unattended agent honest.
2. 16Authoring an MCP Server10 minThe customer-facing boundary of your agent — what the server exposes is what the agent can touch at the customer's site. Tool schemas, transports, and the security review you will lose if you expose the wrong thing.
3. 17Sub-Agents and Skills9 minContext isolation through sub-agents, on-demand loading through skills, and the coordinator-plus-specialists shape of a small team — applied to a real customer-specific data-access skill.
4. 18Agent Teams in Production9 minFan-out, fan-in and the cost curve of running more than one agent at a time. When a human is the right node in the graph, and the failure modes a single agent does not have.

The spine of the course and the most-failed topic in OpenAI FDE interviews. Judge rubrics, dataset curation, regression gates, and the difference between offline and online that most teams get wrong.

1. 19Why Evals Are the Spine9 minThe single investment that separates a POC that ships from a POC that stalls, and why candidates miss it in the OpenAI interview loop until it is the round they lose.
2. 20Judge Rubrics and LLM-as-Judge10 minHow to produce a reproducible score for an open-ended output — the rubric a second person could apply and agree, and the model that can apply it for you without drifting.
3. 21Dataset Curation and Regression Gates10 minWhere an honest eval set actually comes from, the leakage failure that quietly inflates every score, and how to wire the result into a gate that blocks a bad release before it leaves CI.
4. 22Offline vs Online Evals9 minTwo different questions evals can answer, the incident category only one of them catches, and the A/B test as the specific online shape with the statistics everyone pretends they remember.

Auth, data residency, audit; hosted versus VPC versus on-prem; observability and what an incident response actually looks like when the agent is the thing on fire.

1. 23Auth, Data Residency and Audit9 minThe three things a customer's security review will ask about before anything else — and the specific shape of each answer that gets you past it.
2. 24Deployment Shapes — Hosted, VPC and On-Prem9 minThe three shapes an FDE deploys into, the question to ask before promising any of them, and the cost and support story for each.
3. 25Observability and Incident Response10 minWhat to trace, which four numbers to put on the wall, and what an incident actually looks like when the thing on fire is the agent itself.

Customer discovery, scoping a POC that leads to production rather than a demo that doesn't, the Anthropic, OpenAI and Palantir interview loops with the published rubrics, a worked case study, and the capstone exercise.

1. 26Customer Discovery10 minThe discovery call as a scoping instrument rather than a demo opportunity — the questions that surface the real problem, and the discoveries that disqualify an opportunity you should not take.
2. 27Scoping a POC That Leads to Production11 minThe three commitments to insist on before a line of code is written, the exit criteria that let a POC end rather than drift, and the demo trap that turns a successful-looking POC into a cost.
3. 28The Interview Loops at the Top Labs11 minThe Anthropic, OpenAI and Palantir FDE loops as they stood in 2026 — the rounds, the rubrics, the signals interviewers read for, and the rounds candidates most reliably fail.
4. 29A Worked Case Study12 minA full FDE engagement walked end to end — discovery, scope, POC, evals, MCP tool design, deployment shape, a day-14 incident, and the iteration that followed. A mid-market finserv customer automating expense-report compliance review.
5. 30The Capstone10 minBuild a real end-to-end FDE deliverable — an agent plus an MCP server plus an eval suite plus a one-page memo to a fictional customer CTO. The rubric, the shape, and how to turn it into a portfolio piece a hiring manager will actually read.