Forward deployed engineer skills
The skills that separate a forward deployed engineer from a strong staff engineer — and a practical checklist for screening them.
August 17, 2026
The short answer
Forward deployed engineer skills are senior production software engineering plus three capabilities that are slow to train: becoming fluent in a new business domain in days, shipping through security and legal review at a large organization, and accepting accountability for a business result rather than a ticket queue.
If a candidate is elite at distributed systems and has never merged code inside someone else's compliance boundary, they are not yet an FDE. They may become one. Do not staff a six-week production AI embed on potential.
The hiring guide has interview questions. This page is the skill model those questions are testing.
The skill stack
| Layer | What it looks like in the job | Red flag |
|---|---|---|
| Production engineering | Ships features through CI, review, and on-call like a staff IC | Portfolio of notebooks and prototypes only |
| Domain fluency | Credible architecture after two days with a new claims, trading, or clinical workflow | Needs a quarter of onboarding before the first useful PR |
| Organizational navigation | Gets a change through security, legal, and CAB without going around them | Only worked at startups with no review path |
| Outcome ownership | Ties work to a production metric, not story points | Speaks only in tasks completed |
| Stack flexibility | Builds on the customer's AWS, Azure, on-prem, or .NET shop | Insists on importing a personal stack |
| Transfer | Leaves runbooks, evals, and a named internal owner | Knowledge stays in Slack DMs |
Production engineering is the floor, not the differentiator
You still need someone who can design a retrieval pipeline, write tests, and not set production on fire. Typical depth areas:
- Data and integration: warehouses, APIs, identity, object stores
- ML / LLM systems: RAG, eval harnesses, guardrails, tracing
- Cloud and security: VPC, IAM, secrets, audit logs
- Delivery: CI/CD, feature flags, progressive rollout
The RAG vs. fine-tuning guide is an example of the architecture judgment you want on day one. You do not need every FDE to have trained a 70B model. You need them to know when not to.
The three skills Palantir and OpenAI actually screen
Palantir's FDE loop and OpenAI's enterprise loop both assume coding competence, then pressure-test the rest.
1. Customer context switching
The engineer will not get six months to "learn healthcare." They get hours with an operations lead and must ask questions that expose data lineage, compliance, and the real workflow — then propose an architecture that could ship.
2. Organizational navigation
Production AI in a bank, health system, or law firm is a change-management problem. The skill is empathy for the reviewer, a concrete mitigation (not a circumvention), and a merged PR. Candidates who only rant about "slow enterprise IT" will stall your timeline.
3. Outcome ownership
An FDE is not done when the spike is interesting. Done is: running on real traffic, evals in CI, an owner who can operate it. That is closer to a founding engineer than to a staff-plus IC who waits for a PM.
These three are why compensation sits in the $220k–$600k band in-house. They are also why a definition of the role always includes commit rights and standups, not advisory time.
Skills you can teach vs. skills you must hire
Teachable on an embed (weeks): a new cloud, a new orchestration library, a new internal API.
Not teachable on an embed: judgment under incomplete information, political fluency in a regulated org, the habit of writing the runbook before the victory lap.
Hire for the second list. Train the first on the customer's stack — that is the job.
A screening checklist
- Ask for a production system they shipped inside another company's perimeter.
- Ask them to walk a first week on an unfamiliar stack (for example, .NET + Azure + SOAP). Listen for a small PR that builds trust, not a rewrite proposal.
- Give a regulated scenario (privilege, PHI, SOX) and require access control at retrieval time, not "we'll fine-tune on the corpus."
- Ask what they would do if the internal champion left in week six. Owners talk stakeholders. Ticket-takers talk blockers.
- Only then talk compensation.
If you do not have eight weeks to run that loop, an agency embed is a faster way to get the skill set on a single workflow. That trade-off lives in the hiring guide.
FAQ
Does an FDE need a PhD or publications?
No. The role is production engineering in a customer environment. Research depth helps on some workloads; it does not replace organizational navigation.
Is "full-stack" enough?
Full-stack plus none of the three Palantir/OpenAI skills is a staff engineer. You will like them. They will not get your AI workflow through legal.
Should we hire for a specific model vendor?
No. Hire people who will use Bedrock, Azure OpenAI, Vertex, or on-prem weights depending on your constraint. Vendor loyalty is an SE trait, not an FDE one — see FDE vs solutions engineer.
Need these skills in your repo this quarter?
FDE Agency embeds a named engineer who already has this skill set. Six weeks to a production workflow. Fixed fee. Your IP.
Book a 15-minute scoping call