What is Forward Deployed Engineering (FDE)? Why are AI giants like OpenAI and Anthropic pushing FDE, and will it be the next career to transition into?

Why Are AI Companies Fiercely Pursuing FDE?

@vasuman makes this point directly: "If intelligence itself is becoming commoditized, the only competitive advantage lies in 'how to use it and where to apply it.'"

Model capabilities will be leveled by Anthropic, OpenAI https://t.co/LGL5JOUMli" / X

What is Forward Deployed Engineering (FDE)? Why are top AI companies like OpenAI and Anthropic pushing FDE? Will it be the next career shift worth pursuing? Why are AI companies fiercely pursuing FDE?

The logic is straightforward: If intelligence itself is becoming commoditized, the only competitive advantage lies in "how to use it and where to apply it." Model capabilities will be leveled by Anthropic, OpenAI, and others. Repackaged products can also be replicated. What's truly hard to replicate is embedding AI into a specific company's business workflows. This can't be solved with generic products—it requires sending people to do the work.

Thus, applied AI companies' business model is to deploy FDE engineers on-site to act as "AI transformation outsourcing," with clients paying for efficiency gains. Someone who can independently handle "understand client problems → write code in unfamiliar codebases → explain commercial value to non-technical executives" is what vas calls a "million-dollar hire."

Core requirement of the role: Must be On-site!

This borrows from Palantir's tradition (the origin of FDE definition):

• In 2010, Palantir's FDE engineers embedded with U.S. special forces in Afghanistan, coding at night while troops operated during the day.
• Palantir's CTO stated: "You can't build a product for an environment you're not embedded in."

Applied to AI contexts: Real efficiency gains require "rebuilding the company around AI," which can't be done remotely. You must sit with clients to build custom agents using their proprietary data and context.

FDE's Three-Stage Workflow

1. Audit (Audit/Diagnosis): Ends with a prototype demo

• Rotate through client departments (e.g., RevOps for 2 weeks, procurement for 1 week, finance for a month)
• Objectives:
• Map each team's workflows
• Identify bottlenecks
• Determine what to automate vs. leave manual

Three practical criteria for deciding whether to deploy an agent:

• Rules can be abstracted, but inputs are varied (emails/PDFs/scans) requiring tool orchestration? Deploy an agent!
• Both rules and inputs are predictable? Write regular code—it's faster and cheaper!
• Requires pattern recognition + domain expert judgment? Keep it human!

Two additional heuristics:

• Scale matters: A process run only 5 times/month won't justify ROI.
• Don't over-AI everything: Most tasks need "a series of tool calls + LLM orchestration." Overusing AI increases token costs and reduces quality.

1. Evals (Validation)

When clients spend millions on AI deployment, they need proof "it's actually working." Good evaluations don't just check final answers—they verify AI thinks like humans. Two methods:

• Score step-by-step like humans: Break down expert problem-solving into checkpoints and see if the AI passes each stage.
• Anchor to golden samples: Co-create 20 "perfect answers" with senior employees as benchmarks.

Evals' true purpose is to convince AI-skeptical executives to approve—acting as a business trust tool, not just an engineering tool.

1. Deployment (Deployment)

Counterintuitive but pragmatic principles:

• Avoid massive data migrations. Build APIs over existing data layers (SharePoint/databases) to let models act as orchestrators. Clients won't let you redo ERP systems.
• Start with sandbox environments for safe testing within client infrastructure.
• Begin with minimal autonomy, then escalate permissions. Example: Start with "detect bug → investigate → log ticket," then later allow "write code + submit PR."

How to Become an FDE in 30 Days?!

vas says three backgrounds are easiest to transition: consultants, PMs, software engineers.

Consultant/PM Weakness: Engineering capability Solution: Build a portfolio. Pick two projects from:

• A production-grade agent replicating a full workflow from your previous company (API calls, thought logging, error handling).
• A domain-specific RAG pipeline (legal/medical/financial datasets).
• A custom eval framework with multi-dimensional scoring (accuracy, format, cost, latency).
• A MCP (Model Control Protocol) integrating LLMs into legacy non-AI systems.

vas emphasizes: "Do not outsource your understanding to AI"—letting AI do the thinking makes interviews collapse.

Engineer Weakness: Communication Engineers doing similar projects must explain:

• Component choices
• Technical trade-offs
• Iteration processes
• Business outcomes

And answer: "Why solve this pain point? How would this work in real client scenarios?"

30-Day Roadmap (Role-agnostic) Week 1: Agent loop fundamentals (read Anthropic's Building Effective Agents), tool usage, guardrails, context vs external memory, audit trails Week 2: Structured outputs (JSON), pitfalls moving from demo to production, checkpoint mechanisms Week 3: Retry strategies with exponential backoff, cost optimization (small models for small tasks/caching/token limits), golden dataset creation, multi-agent architectures Week 4: Retrospectives + storytelling, tying everything to business metrics

Quote

vas

@vasuman

13h

Forward Deployed Engineering 101

By the end of this article, you should understand why Anthropic, OpenAI, Google, and other AI companies are looking for FDEs, and how you can capitalize on this demand. I've done the job myself, hired...