Rohit Ghumare · August 1, 2026 · 7 min read

What a Forward Deployed Engineer Actually Does

Bob McGrew, who helped build the role at Palantir, defines it in one sentence: an engineer who sits at the customer site and fills the gap between what the product does and what the customer needs. That sentence survives intact in every 2026 posting from OpenAI to Anthropic to the legal-AI startups. What changed is what fills the gap. This is the job from the inside: the deliverables the postings actually name, the team shape, the pay with sources, and the three tests that separate the real thing from a retitled solutions engineer.

The shape of the work

You are dropped into an organization that has bought an AI product and cannot yet extract value from it. Your first weeks are archaeology, not engineering: finding the workflow nobody documented, the data source people actually trust, and the person who knows why the process works the way it does. Then you build in the gap, on site, against their systems. Travel expectations in current postings run 25 percent at Anthropic and up to 50 percent at OpenAI; the Palantir ancestors lived at the customer for months.

The deliverables are concrete, and the postings have gotten refreshingly specific about them:

• Integrations with the customer's actual systems. Harvey's founding-FDE posting names them: document management, knowledge management, ticketing, identity and SSO. This is where most of the hours go.
• Agents built for their workflows. Anthropic's posting lists the artifacts verbatim: MCP servers, sub-agents, and agent skills that will be used in production workflows. Not demos. Production.
• Per-customer evals. Harvey again: build and maintain evals and harnesses that capture real-world quality on a client-by-client basis. A probabilistic system without a customer-specific eval is a liability with a login page.
• Runbooks and playbooks. Training the humans, writing down what works, and handing off something durable when you rotate out.
• Field reports that change the product. OpenAI's posting: codify working patterns into tools, playbooks, or building blocks that others can use, and share field feedback that helps Research and Product understand where the models succeed.

Concretely, the second bullet is a program you write. An MCP server for the customer’s document system is a few hundred lines, and its shape tells you most of what the job values: narrow tools, read-only until proven, everything logged.

// mcp server for the client's DMS: narrow, read-only, audited
const server = new McpServer({ name: "acme-dms", version: "0.1.0" });

server.registerTool("search_documents", {
  description: "Search matters by query and matter number",
  inputSchema: { query: z.string(), matterId: z.string().optional() },
}, async ({ query, matterId }) => {
  audit.log("search", { query, matterId, user: ctx.user });   // every call logged
  return dms.search(query, { scope: matterId, limit: 20 });    // scoped retrieval: week 7's eval win
});

server.registerTool("get_document", {
  description: "Fetch one document body by id",
  inputSchema: { docId: z.string() },
}, async ({ docId }) => dms.fetch(docId, { readOnly: true })); // no write scope exists yet

// write tools arrive only after the eval harness proves the read path

Nothing in that sketch is clever, and that is the point. The value is in what it encodes about the deployment: retrieval scoped the way the client’s matters are actually organized, an audit trail their security team can read, and authority withheld until the proof exists.

The team shape

Nobody deploys alone. Palantir's original pods were four or five engineers, the "Deltas," each pod paired with an "Echo," a deployment strategist who owned the customer relationship and translated the business problem; Echoes were often former military officers, clinicians, or forensic accountants rather than engineers. The modern versions keep the pairing: OpenAI's postings reference deployment strategists and solutions architects alongside FDEs, and AWS runs pods of five to six engineers on 45-day sprints. The pattern to notice: the engineer is protected from being the account manager, so the engineer can stay an engineer.

Not a solutions engineer, not a consultant

Those three rightmost cells are the tests. The a16z history is blunt that the title began as a rebrand of solutions engineering, and the Palantir veterans are equally blunt that most of today's hundreds of FDE-hiring companies are running the rebrand without the design. When you interview, ask all three: will I ship code into the product, what am I measured on six months after go-live, and show me one platform feature that started as a field hack. The answers tell you which job you are actually taking.

Skills, from a thousand postings

The Bloomberry analysis of roughly 1,000 FDE postings in late 2025 gives the distribution: Python required in 66 percent, TypeScript in 35, agent experience named in 35, LLM experience in 31. Top verticals: financial services at 24 percent, government and defense at 18, healthcare at 17. Seventy percent of postings include equity. The requirements that repeat across OpenAI, Anthropic, and the startups: full-stack shipping ability, agent development, eval frameworks, prompt engineering that goes past vibes, and a phrase that appears in some form everywhere, high agency in ambiguity. The role is engineering plus the social courage to walk into someone else's building and start asking why.

The technical stack is exactly the one I keep writing about: the job is harness engineering performed at the customer's elbow. Context, tools, loop, proof. If you can build a per-customer eval harness and wire an agent into a document system behind SSO, you can do the core of this job.

The pay, with sources attached

All US ranges. Third-party reports circulate much higher total-compensation figures for frontier labs; treat anything beyond the posted ranges and levels.fyi aggregates as unverified marketing until you see an offer letter. The honest summary: senior product-engineer money, plus equity in 70 percent of postings, plus a travel schedule you should price in.

Getting in

The credential that works is a deployed system with a named user, because the job is deployed systems with named users. Build an agent that does real work against real infrastructure, integrate it with the boring enterprise surfaces (SSO, a ticketing system, a document store), write the eval harness that proves it works, and write up the deployment including what broke. That artifact answers every FDE interview question at once. The postings ask for three to five years of engineering with customer-facing exposure; the portfolio substitutes for some of it, vertical knowledge (finance, healthcare, legal, defense) substitutes for more. If you are starting from zero on the agent side, my AI engineering path covers the foundations; the FDE role is that skill set pointed outward, at someone else's building.

For why this job exists at all, and why four companies just committed nine billion dollars to it, read the companion essay: The Forward Deployed Engineer. Sources: the OpenAI and Anthropic postings, Pragmatic Engineer's deep dive, Bloomberry's posting analysis, and levels.fyi.