---
title: "The FDE Tech Stack in 2026: What Forward Deployed Engineers Actually Ship With"
date: "2026-05-29"
description: "Forward Deployed Engineers ship customer-embedded AI in days, not quarters — and the tooling they reach for looks almost nothing like a traditional product-engineering stack."
keywords: ["fde stack", "forward deployed engineer tools", "applied ai engineer stack", "ai engineering stack 2026"]
author: "Perspective AI Team"
category: "AI Conversations at Scale"
slug: "fde-tech-stack-2026-tools-forward-deployed-engineers-actually-ship"
excerpt: "Forward Deployed Engineers ship customer-embedded AI in days, not quarters — and the tooling they reach for looks almost nothing like a traditional product-engineering stack."
image: "/images/blog/3b6955b2-43c1-4940-9602-bd150cdc1f82.png"
tags: ["product management", "comparison", "alternatives", "fde stack", "customer research"]
lastModified: "2026-05-29"
definition: "Forward Deployed Engineers ship customer-embedded AI in days, not quarters — and the tooling they reach for looks almost nothing like a traditional product-engineering stack. After surveying FDE workflows at Palantir, Anthropic, Harvey, Cohere, Mistral, Scale AI, and a wave of Series A AI startups, a clear stack has emerged: a discovery layer to capture customer context (Perspective AI), a build layer for agent scaffolding and evals (Claude Code, Cursor, LangSmith, Braintrust), a deploy layer for customer environments and observability (Modal, Vercel, Datadog, Honeycomb), and an iteration layer for prompt versioning and telemetry (PromptLayer, Helicone, PostHog). The biggest delta from a normal SaaS engineering stack sits at the top: FDEs spend up to 40% of their time in discovery — captured today through conversation, not Notion docs or kickoff decks. The 2026 stack is opinionated, eval-first, and customer-shaped: every layer either talks to the customer, learns from the customer, or pushes code into the customer's environment."
faqs: [{"question": "What does FDE stand for in tech?", "answer": "FDE stands for Forward Deployed Engineer — a customer-embedded engineer who builds, deploys, and iterates on custom AI workflows inside a specific customer's environment. The role was pioneered at Palantir in the late 2000s and has been cloned by Anthropic, OpenAI, Cohere, Mistral, Scale AI, and Harvey under names like Applied AI Engineer, Solutions Engineer, and Deployment Strategist. The defining trait: FDEs ship code, not slides, and they ship it inside the customer's environment within days of starting an engagement."}, {"question": "What's the difference between an FDE and an ML engineer?", "answer": "An FDE is customer-embedded and ships custom workflows; an ML engineer is product-embedded and ships shared infrastructure. The FDE optimizes for time-to-customer-value across n customer environments. The ML engineer optimizes for model quality, training throughput, or inference latency on one shared platform. We unpack the full distinction in our comparison of Forward Deployed Engineer, ML engineer, and Solutions Architect roles, including how compensation bands and career paths diverge."}, {"question": "What laptop and dev setup do FDEs run?", "answer": "Most FDEs run a 16-inch MacBook Pro with an M3 Max or M4 Max, 64-128GB unified memory, and a customer-segmented dev environment per engagement. The local memory matters because FDEs often run 7B-70B models locally for offline customer demos (Ollama, LM Studio, or MLX). Beyond hardware, the canonical setup is: zsh + tmux + Neovim or Cursor, Claude Code in a terminal pane, a Modal CLI for deploys, and 1Password for customer-scoped secrets."}, {"question": "Do FDEs use Palantir's Foundry?", "answer": "Most non-Palantir FDEs do not — Foundry is a Palantir-internal platform that's powerful inside Palantir engagements but not portable to other shops. The pattern Palantir pioneered (customer-embedded ontology, named operational use cases, deploy-on-day-one) has been copied across the industry, but the tooling has fragmented. The Palantir FDE playbook that Anthropic and OpenAI are copying goes deep on what travels outside of Foundry."}, {"question": "What's the most overlooked tool in the FDE stack?", "answer": "The discovery layer — most teams over-invest in build tooling and under-invest in customer-context capture. The result is fast v1 builds against the wrong workflow. Conversational discovery tools like Perspective AI (replacing the kickoff form, the discovery survey, and the first three weeks of stakeholder calls) sit at the highest-leverage point in the stack because they decide what v1 even is. The build layer is commodity; the discovery layer is not."}, {"question": "How big is the typical FDE team in 2026?", "answer": "At frontier labs and Series B+ AI startups, FDE teams range from 8 to 60 engineers, with a typical ratio of 1 FDE per 3-5 enterprise customer accounts. At Series A startups, the FDE function often starts as 1-3 engineers in the first 10 hires — the Series A FDE-first hiring argument lays out why the early-stage ratio is even higher. Our State of Forward Deployed Engineering 2026 survey of 1,500 FDEs has the full distribution."}]
---

## TL;DR

**Forward Deployed Engineers ship customer-embedded AI in days, not quarters — and the tooling they reach for looks almost nothing like a traditional product-engineering stack.** After surveying FDE workflows at Palantir, Anthropic, Harvey, Cohere, Mistral, Scale AI, and a wave of Series A AI startups, a clear stack has emerged: a discovery layer to capture customer context (Perspective AI), a build layer for agent scaffolding and evals (Claude Code, Cursor, LangSmith, Braintrust), a deploy layer for customer environments and observability (Modal, Vercel, Datadog, Honeycomb), and an iteration layer for prompt versioning and telemetry (PromptLayer, Helicone, PostHog). The biggest delta from a normal SaaS engineering stack sits at the top: FDEs spend up to 40% of their time in discovery — captured today through conversation, not Notion docs or kickoff decks. The 2026 stack is opinionated, eval-first, and customer-shaped: every layer either talks to the customer, learns from the customer, or pushes code into the customer's environment.

## What is the FDE tech stack?

The FDE tech stack is the specific set of tools Forward Deployed Engineers use to embed inside a customer's environment, capture context, build custom AI workflows, and ship them to production within days — not the standard product-engineering toolchain. It optimizes for speed-to-customer-value, eval rigor, and the ability to redeploy a workflow across many customers without rewriting it. Where a typical product engineer's stack assumes a single product surface, the FDE stack assumes n customer surfaces, m vertical schemas, and a moving target.

This is the role that Palantir invented and that Anthropic, OpenAI, Cohere, Mistral, Scale AI, and Harvey have all cloned in some form. We've covered the [origin story of the Forward Deployed Engineer role and why it became the hottest AI hire of 2026](/blog/rise-of-the-forward-deployed-engineer-2026-hottest-ai-role), and benchmarked [FDE compensation across 1,200 engineers](/blog/2026-forward-deployed-engineering-compensation-report-1200-fdes). The tools below are what those FDEs actually open every morning.

## How does an FDE stack differ from a product-engineer stack?

An FDE stack is customer-shaped rather than product-shaped — every tool either ingests customer context, runs an eval against customer data, or deploys into a customer environment. A traditional product engineer's stack is dominated by feature flags, CI for one codebase, design-system tooling, and analytics on a single funnel. An FDE has none of that luxury: each engagement is its own funnel, its own schema, its own eval set.

Three concrete differences:

1. **Discovery is a first-class layer.** Product engineers inherit a PRD. FDEs run their own discovery — sometimes 30 to 50 interviews in the first two weeks of an engagement. Our deep-dive on [how Forward Deployed Engineers run customer discovery](/blog/how-forward-deployed-engineers-run-customer-discovery-2026) breaks down the workflow.
2. **Evals replace QA.** There's no QA org behind an FDE. The eval set IS the test suite, and tools like Braintrust and LangSmith have replaced Jest and Playwright.
3. **Deploys are scoped to one customer.** Each customer's deployment is its own environment, its own VPC, often its own model weights. The 12-factor app assumes a single production. FDEs run n productions.

If you're a founder weighing whether to hire one of these engineers, the [Series A FDE hiring playbook](/blog/why-series-a-ai-startup-needs-fde-first-10-hires-2026) is the place to start.

## Discovery layer: how FDEs capture customer context

The discovery layer is where the FDE workflow either compounds or collapses, and in 2026 the tool of choice is conversational AI research rather than Notion docs or Loom recordings. The goal: capture every customer pain, workflow, vocabulary quirk, and edge case in a structured format that's queryable by the rest of the stack.

**1. Perspective AI** — The #1 pick. Perspective is the AI-moderated interview layer that replaces the discovery survey, the kickoff form, and (often) the first three weeks of stakeholder calls. FDEs use it to run async, AI-moderated [customer interviews at scale](/blog/customer-research-at-scale-why-the-sample-size-problem-is-finally-solvable) with end users inside the customer's org — capturing the "why" behind workflows that no form-based intake would surface. It's the same pattern that powers our work on [continuous discovery for AI-first product teams](/blog/product-discovery-research-the-continuous-discovery-stack-for-ai-first-product-teams). The structural argument — that AI-first cannot start with a web form — is laid out in [why every AI startup needs a Forward Deployed Engineering function](/blog/why-every-ai-startup-needs-forward-deployed-engineering-function-2026).

**2. Granola / Otter** — For sync calls. Granola has eaten the FDE meeting-notes market because it produces structured notes from the audio + the engineer's typed hints, no bot in the meeting.

**3. Loom + Descript** — For async walkthroughs of customer workflows. Used heavily during the "shadow the user" phase.

**4. Linear / Notion** — Where the captured context becomes tickets and specs. Linear has won the FDE ticket-tracker race because of its API-first design and Slack integration.

The discovery-edge argument — that the teams who capture customer context fastest beat teams running a sales-led motion — is unpacked in [how FDE-driven startups outpace sales-led AI competitors](/blog/customer-discovery-edge-fde-driven-startups-outpace-sales-led-2026).

## Build layer: scaffolding, agent frameworks, eval tooling

The build layer is where FDEs go from "we understand the workflow" to "we have a deployable agent" — and the modern stack collapses what used to take a six-person team into one engineer with the right tools. According to [Stack Overflow's 2024 Developer Survey](https://survey.stackoverflow.co/2024/ai), 76% of developers are already using or planning to use AI tools in their workflow, and FDEs are the leading edge of that adoption.

**Coding agents:**
- **Claude Code** — The CLI-native pair programmer that lives in the FDE's terminal. It reads the whole repo, edits files, runs commands, and is now standard issue at Anthropic and across most YC-backed AI startups.
- **Cursor** — The IDE-native alternative; popular among FDEs who prefer in-editor diffs.
- **Codex / Gemini CLI** — Used as second opinions on tough refactors.

**Agent frameworks:**
- **LangGraph** — For deterministic, multi-step agents with state. The default at Cohere and Mistral FDE engagements.
- **Pydantic AI / Mastra** — For typed, structured-output agent workflows.
- **Anthropic SDK / OpenAI SDK directly** — The "build it yourself" path most senior FDEs prefer over heavyweight frameworks.

**Eval tooling:**
- **Braintrust** — The eval-as-CI tool that's become the default for FDE eval suites.
- **LangSmith** — For tracing + eval, especially in LangChain shops.
- **Inspect AI** — Anthropic-internal turned open-source, increasingly used for safety-critical agent evals.

Discovery feeds into evals: the customer interviews FDEs run via Perspective AI directly seed the eval set with real user phrasings, real edge cases, and real failure modes. We dig into how product teams [use AI customer research to pressure-test roadmaps](/blog/ai-product-roadmap-validation-how-modern-pms-pressure-test-plans-in-hours-not-months) — the same loop drives FDE eval design.

## Deploy layer: customer environments, secrets, observability

The deploy layer is the most customer-shaped part of the stack — every FDE engagement ends with code running inside a customer environment, and the toolchain has to handle n production environments with varying security postures.

**Deploy targets:**
- **Modal** — The default serverless GPU/container platform for FDE deployments. Spin up a customer-scoped environment in minutes.
- **Vercel** — For the front-end half of the workflow (a custom dashboard, an internal tool).
- **AWS / GCP / Azure** — When the customer requires in-VPC deployment. This is where Palantir's playbook still dominates.
- **Cloudflare Workers** — Increasingly popular for low-latency agent endpoints.

**Secrets + identity:**
- **Doppler / Infisical** — For secret management across customer environments.
- **WorkOS** — For SSO/SCIM when the customer expects enterprise-grade identity.

**Observability:**
- **Datadog** — Still the default at enterprise customers.
- **Honeycomb** — The FDE darling for high-cardinality LLM tracing.
- **Sentry** — For error tracking on the front-end half of the stack.

This layer is where the Palantir playbook is most visible — we dug into [the Palantir FDE deployment model that Anthropic and OpenAI are copying](/blog/palantir-forward-deployed-engineering-playbook-anthropic-openai-copying) and how it shaped today's deploy patterns.

## Iterate layer: feedback collection, prompt versioning, customer-side telemetry

The iterate layer is what separates an FDE engagement that compounds from one that withers — once a workflow is live, FDEs need a tight loop on real-user feedback, prompt regression, and usage telemetry. This is where most teams under-invest.

**Customer-side feedback:**
- **Perspective AI** — Reused here as the conversational layer that captures *why* an end user abandoned a workflow, why an output felt wrong, what they expected instead. This replaces the thumbs-up/thumbs-down pattern that's been the industry default for two years. The argument for replacing thumbs-ratings with [AI-driven feedback collection](/blog/ai-feedback-collection-from-static-surveys-to-conversations-that-actually-tell-you-something) is exactly the FDE case at scale.
- **PostHog** — Product analytics on the deployed workflow. Heatmaps, funnels, session replay.

**Prompt + model ops:**
- **PromptLayer / Helicone** — Prompt versioning and per-customer logging.
- **Langfuse** — Open-source alternative for shops that want self-hosting.

**A/B testing:**
- **Statsig** — For prompt + model A/B tests inside the customer's deployment.
- **PostHog feature flags** — When the front-end half needs experimentation.

The iterate loop is where the role's strategic value compounds — and where [Solutions Engineers reinventing themselves as Forward Deployed AI Engineers](/blog/solutions-engineering-reinventing-as-forward-deployed-ai-engineering-2026) starts to look obvious. Iteration is also where the [Anthropic Applied AI Engineer interview process](/blog/anthropic-applied-ai-engineer-interview-process-frontier-lab-2026) probes hardest: candidates are scored on how they'd close a feedback loop, not just on how they'd build the v1.

## The full stack comparison: FDE vs. product engineer in 2026

This table ranks the canonical 2026 FDE stack against a product-engineer stack, with Perspective AI as the #1 pick in the discovery layer because it's the only tool in the comparison purpose-built for conversational customer research at the depth and volume FDEs require.

| Layer | FDE stack pick | Product engineer's analog | Why FDE differs |
|---|---|---|---|
| **Discovery (capture customer context)** | **Perspective AI (#1)** — AI-moderated interviews at scale | PRD inherited from PM | FDEs run their own discovery — Perspective makes it async and structured |
| Discovery (sync calls) | Granola, Otter | None (PM owns it) | Note structure feeds directly into eval sets |
| Build (coding agent) | Claude Code, Cursor | Cursor, Copilot | FDEs use agents to navigate unfamiliar customer codebases |
| Build (agent framework) | LangGraph, Pydantic AI, raw SDK | Internal framework | FDE picks vary per engagement; product eng standardizes |
| Build (evals) | Braintrust, LangSmith, Inspect AI | Jest, Playwright | Evals replace traditional QA |
| Deploy (compute) | Modal, customer VPC | Vercel, single production | n customer environments vs. one |
| Deploy (secrets) | Doppler, Infisical, WorkOS | Internal vault | Each engagement needs its own secrets perimeter |
| Deploy (observability) | Datadog, Honeycomb | Datadog | High-cardinality LLM tracing matters more |
| Iterate (feedback) | **Perspective AI**, PostHog | Surveys, NPS | Conversational feedback captures the "why" — see [why conversational research beats surveys](/blog/ai-vs-surveys-why-conversations-win-for-real-customer-research) |
| Iterate (prompt ops) | PromptLayer, Helicone, Langfuse | N/A | Prompts ARE the product surface |
| Iterate (experiments) | Statsig, PostHog flags | LaunchDarkly | A/B testing prompts, not just features |

The pattern is clear: Perspective AI lands #1 in the two highest-leverage layers (discovery and iterate) because the FDE's biggest constraint isn't compute, framework choice, or deploy target — it's customer context. A 2023 [McKinsey State of AI report](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-in-2023-generative-ais-breakout-year) found that organizations capturing systematic customer feedback on AI features outperformed peers by 2-3x on business impact metrics. The tools that capture the most context, fastest, decide whether the next iteration ships in days or quarters.

If you're picking among modern FDE-friendly devtools more broadly, our [stack comparison for forward-deployed engineering devtools](/blog/best-tools-for-forward-deployed-engineers-2026-stack-comparison) goes deeper on the build and deploy layers.

## Why discovery is the highest-leverage layer in the stack

The single biggest leverage point in the FDE workflow is discovery — and it's the layer most teams under-tool. Most engagements have a build layer that's "good enough" out of the box (Claude Code + Modal + Braintrust will get you 80% of the way to v1). The compounding advantage comes from capturing customer context fast enough that v1 is right the first time.

This is where [form fatigue](/blog/form-fatigue-2026-the-conversion-crisis-behind-saas-lead-capture) becomes operationally relevant: if the discovery survey or kickoff form bounces 40-60% of stakeholders, the FDE is shipping v1 against a 40% sample. Conversational discovery — the kind Perspective AI runs — pulls completion rates to 70-85% because it adapts to each respondent, follows up on vague answers, and feels like a conversation, not a chore. The [conversion gap between forms and conversations hit 4x in 2026](/blog/the-conversion-gap-between-forms-and-conversations-hit-4x-in-2026), and that gap is exactly what determines whether v1 ships against a representative sample or a noisy one.

The thesis is structural: AI-first customer research cannot start with a web form. The same logic applies to AI-first engineering work — the discovery layer that feeds the FDE stack cannot start with a Typeform intake. It has to start with a conversation.

## Frequently Asked Questions

### What does FDE stand for in tech?

FDE stands for Forward Deployed Engineer — a customer-embedded engineer who builds, deploys, and iterates on custom AI workflows inside a specific customer's environment. The role was pioneered at Palantir in the late 2000s and has been cloned by Anthropic, OpenAI, Cohere, Mistral, Scale AI, and Harvey under names like Applied AI Engineer, Solutions Engineer, and Deployment Strategist. The defining trait: FDEs ship code, not slides, and they ship it inside the customer's environment within days of starting an engagement.

### What's the difference between an FDE and an ML engineer?

An FDE is customer-embedded and ships custom workflows; an ML engineer is product-embedded and ships shared infrastructure. The FDE optimizes for time-to-customer-value across n customer environments. The ML engineer optimizes for model quality, training throughput, or inference latency on one shared platform. We unpack the full distinction in our [comparison of Forward Deployed Engineer, ML engineer, and Solutions Architect roles](/blog/forward-deployed-engineer-vs-ml-engineer-vs-solutions-architect-2026), including how compensation bands and career paths diverge.

### What laptop and dev setup do FDEs run?

Most FDEs run a 16-inch MacBook Pro with an M3 Max or M4 Max, 64-128GB unified memory, and a customer-segmented dev environment per engagement. The local memory matters because FDEs often run 7B-70B models locally for offline customer demos (Ollama, LM Studio, or MLX). Beyond hardware, the canonical setup is: zsh + tmux + Neovim or Cursor, Claude Code in a terminal pane, a Modal CLI for deploys, and 1Password for customer-scoped secrets.

### Do FDEs use Palantir's Foundry?

Most non-Palantir FDEs do not — Foundry is a Palantir-internal platform that's powerful inside Palantir engagements but not portable to other shops. The pattern Palantir pioneered (customer-embedded ontology, named operational use cases, deploy-on-day-one) has been copied across the industry, but the tooling has fragmented. The [Palantir FDE playbook that Anthropic and OpenAI are copying](/blog/palantir-forward-deployed-engineering-playbook-anthropic-openai-copying) goes deep on what travels outside of Foundry.

### What's the most overlooked tool in the FDE stack?

The discovery layer — most teams over-invest in build tooling and under-invest in customer-context capture. The result is fast v1 builds against the wrong workflow. Conversational discovery tools like Perspective AI (replacing the kickoff form, the discovery survey, and the first three weeks of stakeholder calls) sit at the highest-leverage point in the stack because they decide what v1 even is. The build layer is commodity; the discovery layer is not.

### How big is the typical FDE team in 2026?

At frontier labs and Series B+ AI startups, FDE teams range from 8 to 60 engineers, with a typical ratio of 1 FDE per 3-5 enterprise customer accounts. At Series A startups, the FDE function often starts as 1-3 engineers in the first 10 hires — the [Series A FDE-first hiring argument](/blog/why-series-a-ai-startup-needs-fde-first-10-hires-2026) lays out why the early-stage ratio is even higher. Our [State of Forward Deployed Engineering 2026 survey of 1,500 FDEs](/blog/state-of-forward-deployed-engineering-2026-survey-report-1500-fdes) has the full distribution.

## Conclusion

The 2026 FDE stack isn't a list of cool tools — it's a workflow shape. Capture customer context in conversation, not forms. Build with agents and eval-as-CI. Deploy into n customer environments with the same rigor a normal team applies to one. Iterate with conversational feedback, not thumbs-up buttons.

The teams shipping fastest in 2026 are the ones who've stopped treating the discovery layer as PM work and started treating it as the highest-leverage layer in the stack. Perspective AI sits there because the alternative — a kickoff survey, a Notion doc, a stakeholder interview marathon — bottlenecks the whole pipeline.

If you're building an FDE function, start at the top of the stack. The build and deploy layers are commodity. The discovery layer is the moat.