The forward deployed engineer handbook

What is a forward deployed engineer?

A customer-facing software engineer who develops and deploys software inside a client organization's own operational environment — post-sale, embedded, staying until it works.

The name borrows the military's “forward deployment” — forces stationed at the point of action, not a rear base. Palantir applied it to engineers around 2010, called the team “Delta” internally, and by roughly 2016 employed more forward deployed engineers than traditional software engineers, per The Pragmatic Engineer. Day to day the work is technical discovery in the client's stack, production code in the customer's repos, data integration, and a constant feedback loop to the product team — with 20–50% travel.

Why every AI company is hiring for it now: enterprise AI's bottleneck moved from model capability to deployment. a16z called it “the hottest job in startups” in 2025, and by 2026 OpenAI had stood up a dedicated deployment business while AWS committed a billion dollars to embedding engineers with customers.

1. 01EmbedSit with the customer's team — often on their site, often in their Slack — until you understand the workflow well enough to argue about it.
2. 02DiscoverMap the messy reality: the systems of record, the data that is wrong, the process nobody documented. This is the round FDE interviews test.
3. 03BuildWrite production code in the customer's stack and repos, against their data and their constraints — not a demo, and not a proof of concept.
4. 04DeployShip it into their environment, through their review, their compliance, and their release process. Owning the outcome means owning this part.
5. 05Feed backTurn what the field taught you into product. This is the line between an FDE and a consultant: the learnings become platform features, not a report.

• Python66%
• TypeScript35%
• AI agents35%
• AWS32%
• LLM experience31%
• RAG12%

Sixty percent of roles want 3–5 years' experience, and 58% of postings come from 11–200-person companies. Interview loops center on a decomposition round — take an ambiguous, messy problem and structure it out loud. Narrated reasoning beats polished answers.

Forward deployed engineer salary, 2026

Glassdoor

~$156K average · $125K–$198K (25th–75th percentile)

Bloomberry (1,000+ postings)

$173,816 median · 70% mention equity · 0% carry quota

Frontier AI labs

$350K–$450K total compensation

All figures approximate · self-reported or posting-derived

Three to five parallel engagements is the normal load, and each one is genuinely separate — its own repository, its own environment, its own stakeholders, its own half-loaded context in your head. Nothing about that composes. Switching from a customer's data pipeline to another customer's auth migration is a full context reload, and an FDE does it several times a day.

Coding agents changed which half of that is expensive. Generating a plausible implementation is now cheap and nearly instant, so the work that remains is running it, reading the diff, deciding whether it is right, and getting it through the customer's review. Agents produce that reviewable output faster than one person can consume it — which means an FDE's throughput is now bounded by review attention, not typing speed. Adding a second agent to the same terminal doesn't help; it just deepens the queue behind one prompt.

So the honest description of the modern FDE bottleneck is not “writing code is slow”. It is that a serial, one-terminal-at-a-time work surface is being asked to carry a parallel workload — and forward deployed engineers hit that wall earlier and harder than almost anyone else in software, because parallel engagements are the job description rather than an unlucky sprint.

Build

A coding agent — Claude Code, Codex, or Cursor's agents — plus an eval harness, so a change is provably better rather than merely different.

Deploy

The customer's own infrastructure most of the time; Modal, Vercel, or Fly when the FDE owns the target. Their compliance, their release process.

Iterate

Telemetry and observability from the deployed system, feeding the next change. Without it an engagement is guesswork after week two.

Work surfacethe gap

Where every engagement is tracked, reviewed, and shipped — in parallel, without a terminal per client. This is the layer nobody built.

Call that fourth layer the engagement board: one place where every client engagement is a lane, every unit of work is a ticket with its own branch and its own agent, and the human's job is to review and ship rather than to remember which terminal was which. Ticket trackers model the work but never run it; editors and terminals run one thread but never hold the portfolio. The engagement board is the missing middle.

One project per client. Each engagement is its own project pointed at that customer's repository, so switching clients is switching boards, not rebuilding context. One ticket, one worker: every ticket gets its own git branch, worktree, and AI agent session, which is what makes four engagements genuinely concurrent instead of four things queued behind one terminal. Work lands in a review column, and the human's job there is the glance — read the diff, ship it or send it back.

Two details matter more for forward deployed engineers than for anyone else. First, it is agent-agnostic and runs on the subscription you already pay for — Claude Code, Codex, or your own shell, whichever one a given client's security review cleared, on your existing plan rather than a metered API bill. Second, it is a desktop app and nothing about a client's code or your agent conversations leaves your machine unless you explicitly turn that on — which is the answer to the question every customer security review eventually asks.

If you want the mechanics, see how the board works. For where it sits among the alternatives, there is the wider AI coding tool landscape, plus head-to-heads against Cursor, Devin, and Linear. The reasoning behind the one-ticket-one-worker model is in spawning agents was never the hard part.

Frequently asked

Which companies hire forward deployed engineers?

Palantir invented the role around 2010. Today OpenAI, Anthropic, AWS, Google, Ramp, and hundreds of AI startups hire FDEs — most postings come from companies of 11–200 people.

Is a forward deployed engineer just a consultant?

No. Consultancies deliver to spec and leave when the engagement ends. FDEs apply the product model to embedded work: what the field teaches becomes platform features, not a report.

Is FDE a good career move in 2026?

Postings grew roughly 1,165% year over year, OpenAI and AWS funded FDE programs with billions of dollars, and median posting pay clears $170K (approximate). It is the clearest fast lane between software engineering and the AI deployment wave.

How much do FDEs travel?

Typically 20–50% of the time — the “forward deployed” part is literal. Some AI-lab roles are remote-first with periodic customer on-sites.

What is the difference between a forward deployed engineer and a sales engineer?

A sales engineer works the pre-sale: demos, proofs of concept, and technical objection handling, usually against a quota. An FDE picks up after the contract is signed and writes production code in the customer's stack. Roughly 0% of FDE postings carry a sales quota, which is the cleanest way to tell the two roles apart in a job description.

What tools do forward deployed engineers use?

A coding agent (Claude Code, Codex, or Cursor's agents) plus an eval harness for the build loop; the customer's own cloud — or Modal, Vercel, or Fly when the FDE owns the target — for deploy; and telemetry from the deployed system to drive the next iteration. The missing layer is the work surface that tracks all of it across several engagements at once.

How many customer engagements does one FDE run at once?

Three to five is typical, and that is the real constraint on the role — not how fast anyone writes code. Each engagement is a separate repo, environment, and stakeholder, so the scarce resource is the FDE's review attention, which coding agents consume faster than they free.