Terms you'll hear in FDE interviews and engagements. Each entry cites the company or source where the term originated; this is the actual industry vocabulary, not made-up jargon.
An engineer whose primary workspace is the customer's environment, not the company's product backlog. Invented at Palantir in 2006 by Shyam Sankar (employee #13, now CTO). Now standard at OpenAI, Anthropic, Ramp, Google Cloud, and others.
Palantir's internal name for the FDE function. From the engineering blog post 'Dev versus Delta': Deltas ship many capabilities to one customer (the deployment).
Contrast with Dev.
Palantir's internal name for the platform engineering function. From 'Dev versus Delta': Devs ship one capability to many customers (the platform). Different optimization function from a Delta.
OpenAI's 2025-announced FDE structure: a $4 billion joint venture with TPG, Advent International, Bain Capital, and Brookfield. Runs OpenAI's FDE motion at enterprise scale. By mid-2025 it had 10+ FDEs across 8 cities, led by Colin Jarvis.
A standardized server interface that lets an AI system talk to a customer's tools. Listed as a primary FDE deliverable in Anthropic's 2026 'Forward Deployed Engineer, Applied AI' job posting. FDEs typically write one MCP server per customer-specific tool integration.
A scoped, autonomous unit that performs a defined task and returns a result. Fits cleanly inside a customer's existing job-queue infrastructure. Named in Anthropic's FDE job specs as a production artifact FDEs ship.
A reusable capability that an AI agent can compose. Named in Anthropic's FDE job specs alongside MCP servers and sub-agents as primary FDE deliverables.
A labeled set of inputs with expected (or graded) outputs that you run before and after every prompt or model change. The AI-era equivalent of unit tests. FDEs spend significant time building customer-specific eval suites because the customer's data is the eval set.
The staged adoption pattern for AI features: suggestion mode (model proposes, human reviews) → supervised autonomy (model acts for safe categories, escalates risky ones) → full autonomy (model just does it). FDEs typically deploy one rung below where the customer says they want to be.
OpenAI's publicly described FDE engagement model: (1) Scoping - a couple of days onsite with the customer, (2) Validation - building evals and quality checks, (3) Delivery - onsite typically a few days per week. Each phase de-risks a specific failure mode.
One user-facing capability implemented through every layer of the stack at its minimum viable version, top to bottom. The FDE alternative to horizontal (per-layer) progress, which is invisible to a customer until every layer finishes.
Don't generalize a Delta's work after the first customer. Don't generalize after the second. The third customer asking for the same capability is the signal that there's a pattern worth productizing.
Shyam Sankar's framing for FDE hiring: 'I don't want to assemble a polite roster of cross-functional professionals. I want the X-Men, a medley of mutants united for good.' Selecting for range and density over polish.
Anthropic's public compensation stance: comp is set by level, not by negotiation. Dario Amodei, Fortune (Aug 2025): 'We don't negotiate that level because we think it's unfair.' Applies to FDE roles too.
The data work that dominates most FDE engagements: integrations, data cleaning, schema reconciliation. Per Pragmatic Engineer reporting, plumbing is usually the majority of FDE engineering effort, not a precursor to it.