Where the role comes from
Palantir sent roughly 120 forward-deployed engineers to embed with JPMorgan in 2009, and the model stuck: engineers who write production code sitting next to the people who use it, instead of a spec passed between teams. (source) OpenAI has since adopted the same shape. Its FDE team under Colin Jarvis grew from 2 to 39 engineers in 2024 and now targets problems worth tens of millions to low billions of dollars, and in May 2026 OpenAI launched a majority-owned Deployment Company (via the Tomoro acquisition, roughly 150 FDEs and deployment specialists) with adopters including HP, Intuit, Oracle, State Farm, and Uber. The reason the role exists: an estimated 95% of enterprise AI deployments fail to reach production, and that gap is what FDEs are hired to close.
The FDE method: audit, evals, deployment
Every credible account of the role, including OpenAI's own case studies, describes the same three-phase loop. Each phase earns the right to the next, then the whole thing repeats.
1. Audit
Learn how the work really happens, not the documented process. Sit with the people doing it. Map every step, exception, and handoff into a before/after operating map.
2. Evals
Turn non-determinism into evidence. Build a golden dataset of real queries with hand-labeled ideal outputs, then track pass rate, failure categories, and escalation. This is how trust gets earned.
3. Deployment
Ship inside the systems that already exist, never a forced migration. Start in a sandbox, increase autonomy gradually, and monitor everything.
This loop is credited to the FDE method as practiced at Palantir and OpenAI, and it has been codified in detail by Vas at Varick Agents' FDE-in-30-Days guide, which is the most complete public walkthrough of the role we have found and the source we point our own team to first.
The 30-day shape
You do not need 30 days to understand the shape, only to practice it once end to end. In broad strokes, spread across four weeks:
- Week 1, build: get a working agent up: looping, tool use, guardrails, context and memory, and a full audit trail of every prompt, tool call, and error. If the client cannot see what the system did, they will not trust it.
- Week 2, production-ready: structured, schema-validated outputs, explicit handling for missing data, malformed responses, dead APIs, timeouts and partial results, and checkpointing so a failure does not mean starting over.
- Week 3, measure and optimize: a golden dataset of 20 or more real cases, evals for correctness, format, tool selection and escalation, cost routing so cheap models handle simple subtasks, and multi-agent only when it genuinely earns its complexity.
- Week 4, communicate and defend: document the pain point, why AI was the right call, the architecture, and what earlier versions got wrong. Pitch it as an engineer (architecture, tradeoffs) and as a stakeholder (problem, outcome, evidence, risk).
This is our summary of the shape, in our own words. For the full day-by-day breakdown, read Varick Agents' FDE-in-30-Days guide.
“There's only one way something can go right. But there's a thousand ways it can go wrong. If you're only building for the way it goes right, you're worth nothing.”
Vas, Varick Agents, FDE-in-30-Days
FDE vs. consultant
The clearest distinction, backed by OpenAI's own account of how its FDE team operates: consultants hand off a recommendation or a spec. FDEs write the production code themselves, embed on-site for weeks (often 4 or more months in regulated industries before earning full trust), use evaluation-driven development instead of opinion, and are kept deliberately small rather than scaled as a billable-headcount practice. The team is judged on a measurable business metric, not hours logged.
On compensation: we are not going to invent a number. Several low-authority sites publish specific FDE salary figures with no verifiable sourcing, and we will not repeat them. Varick's guide states a range of $150K to $1M; that is Vas's claim, not ours, and we are citing it rather than asserting it.
The one principle that matters most
“Do the job before you have the title.” Proof beats credentials. The fastest way to become a forward deployed engineer is to run the audit, evals, deployment loop on a real problem, even a small one, and show the before-and-after.