Anthropic's FDE Job Ad Is a Study Plan. Read It That Way.
The Forward Deployed Engineer role is the hottest title in AI hiring, and most explainers describe the vibe. The actual posting names three deliverables — MCP servers, sub-agents, agent skills. That is not a vibe. That is a checklist you can work through this month.
- #career
- #claude-code
- #agents
Most writing about the Forward Deployed Engineer role explains the shape of it: embedded with the customer, part engineer, part solutions architect, owns the outcome, lots of travel. All true, all useless for actually getting one.
Anthropic's own posting is more specific than any explainer, and it's public. Under what you'd deliver, it names three things: MCP servers, sub-agents, and agent skills, shipped as production artifacts inside customer systems.
Read that as a syllabus. Three named technologies, all documented, all buildable on your own machine tonight.
Why this role exists at all
The pattern is Palantir's, from the mid-2000s: put strong engineers physically inside the customer, build the bespoke thing, then fold what you learned back into the product. The AI labs have adopted it wholesale because frontier models have the same problem Palantir's platform had — enormous general capability, and a gap between that and any particular company's messy systems that documentation cannot close.
Anthropic's version sits in the Applied AI team, embeds with strategic customers, and — this is the part I'd underline — asks you to identify and codify repeatable deployment patterns and feed them back to Product and Engineering. That's the Palantir loop restated. You are not just shipping for one customer; you're the sensor that turns one deployment into a product feature.
The three deliverables, and what each one actually means
MCP servers. The Model Context Protocol is the open standard for connecting a model to tools, databases and APIs. In practice an FDE writes these because the customer's important system — the internal ticket tracker, the proprietary datastore — has no connector and never will. Adding one is a line:
claude mcp add --transport http notion https://mcp.notion.com/mcp
Building one is the real skill. Start by connecting an existing server, then write a small one against an API you already know.
Sub-agents. Specialised workers with their own context window, own system prompt, own tool access. A file in .claude/agents/, name and description required, tools to fence them in. The reason this matters commercially: it's how you stop a customer's long-running job from flooding one context window, and how you enforce that the agent touching production can only read.
Agent skills. The packaged, reusable capability — the thing that turns "I got Claude to do this once" into something the customer's whole team runs the same way.
What the market looks like
Postings for this title have grown steeply across the labs since 2024 — Anthropic, OpenAI, Google, Databricks and Cohere all staff versions of it — and compensation sits well above standard senior engineering bands, with reported ranges commonly starting around $220–280K in the US and going considerably higher at senior levels. Treat any single figure you read as approximate; the bands move fast and vary by level and location.
The trade is travel and ambiguity. Anthropic's posting puts customer-site travel at around 25%, which is the lower end for this category — some competing FDE roles state up to 50%.
Why this is a real opening for a Brazilian dev
Two reasons, and they're structural rather than motivational.
First, the qualifying skills are public and self-teachable. MCP, sub-agents and skills are all documented; nobody needs to give you access to learn them. That is unusual — most high-paying specialisations require being inside a company that already does the thing.
Second, the role selects for something we have a lot of: engineers who are good at landing in an unfamiliar, badly documented codebase and shipping anyway. If you've done consultancy, agency work, or been the person parachuted onto a struggling project, that is the core competency, and it is harder to teach than the tooling.
The honest catch is that these roles are customer-facing and mostly hire in the US and Europe, often with on-site expectations. Build the portfolio anyway — the same three skills are what enterprises hiring directly will want, and that market is far less location-bound.
Do this today
Do not read another explainer. Open Anthropic's actual posting and copy the deliverables into a list.
Then take the smallest one and ship it this week: write one MCP server against an API you already understand, wire it into Claude Code, and use it for something real. That single artifact demonstrates the exact thing the posting asks for — and you'll know more about the role from building it than from any amount of reading about it.
Sources
Keep reading
More guides like this one.
Your AI Fluency Collapses Exactly Where the Code Ships
Anthropic classified 9,830 Claude conversations. The ones that produce code get briefed better and audited worse — fact-checking falls 3.7pp. The index publishes no personal score, so here is how to run one on yourself.
The CV Prompt That Screens You Out First
88% of employers say qualified high-skills candidates get vetted out for not matching the exact criteria in the posting. So don't ask AI to improve your CV — make it screen you out. The prompt, and the three decisions that keep it honest.
The Mock Interview Prompt That Refuses to Help You
The structured interview is the strongest single predictor of job performance, and also the one with the widest spread, because it can be built well or badly. A chat model builds you the bad one. Here is the prompt that installs the structure instead.