# Forward deployed engineering

Forward deployed engineering is best understood through responsibility: taking a customer problem through discovery, implementation, production use, and an operational handoff. The role combines software engineering with domain learning and customer collaboration. Titles, departments, travel expectations, and commercial incentives vary enough that a job title is an unreliable shortcut.

This guide is for engineers worldwide who are considering the role or building evidence that they can do it. Its research cut-off is 14 September 2026. It combines official company announcements, specific employer requisitions, implementation guidance, and explicitly labelled editorial recommendations. It is a curated reference rather than a census of every vacancy or a forecast of employment.

## Findings that change a career decision

The strongest current evidence is specific to an employer and requisition. OpenAI advertises both FDE and forward deployed software engineering positions. Anthropic describes production AI artifacts and customer deployment work. Databricks explicitly places AI FDE delivery within professional services. These distinctions support a functional assessment of work instead of a rigid division between “real FDE” and “consulting.” [OpenAI FDE](https://openai.com/careers/forward-deployed-engineer-%28fde%29-sf-san-francisco/), [Anthropic FDE](https://job-boards.greenhouse.io/anthropic/jobs/5302966008), [Databricks AI FDE](https://databricks.com/company/careers/open-positions/job?gh_jid=8546367002).

Production experience is valuable, but the career is not uniformly senior-only. The existence of a dedicated graduate requisition and an experienced-only requisition is more useful than an undifferentiated statement about average experience. Apply to explicit eligibility criteria and show the work those criteria are intended to capture. [Palantir graduate FDSE](https://jobs.lever.co/palantir/2e6b0ac8-83e9-4be5-a3aa-cf319f751728), [Databricks AI FDE](https://databricks.com/company/careers/open-positions/job?gh_jid=8546367002).

Company investment supports the view that deployment is an organisational priority. It does not establish the number of vacancies or the eventual return on that investment. The guide therefore uses official announcements instead of adding incompatible job-board growth percentages. [OpenAI DeployCo announcement](https://openai.com/index/openai-launches-the-deployment-company/), [Anthropic services-company announcement](https://www.anthropic.com/news/enterprise-ai-services-company), [AWS FDE investment](https://aws.amazon.com/govcloud-us/newsletter/newsletter-issue-04/).

## How to assess a role

Before an interview, write a short account of what you think the role owns. Use six questions:

1. What customer workflow would I improve, and who uses the result?
2. What code or system would I personally build and maintain?
3. How are success and failure measured after launch?
4. Who owns commercial decisions, technical decisions, and incident response?
5. What happens when the engagement ends?
6. Which parts of the work become reusable software, knowledge, or product improvements?

These are editorial due-diligence questions. A good answer names concrete work and boundaries. An answer consisting only of “strategic customers,” “high agency,” or “AI transformation” still leaves the job unresolved.

Ask about the operating conditions separately. Travel percentages do not explain notice periods, overnight stays, weekend travel, on-call expectations, or the number of simultaneous accounts. A remote designation does not establish country eligibility. A sales department does not establish the absence of production work. Salary boilerplate mentioning commission may not describe the incentive plan for the advertised position.

## Building credible evidence

A portfolio should make your decisions inspectable. Choose a workflow small enough to implement end to end, and include these artifacts:

| Artifact | What a reviewer should be able to determine |
|---|---|
| Discovery brief | Who needs the result, the current baseline, and what success means |
| Data and access map | Where the data comes from, who can use it, and what is excluded |
| Runnable implementation | How to reproduce the workflow and where state is stored |
| Evaluation evidence | Which cases were tested, what passed, and what remains unreliable |
| Operating plan | How failures are detected, retried, escalated, and rolled back |
| User feedback | What changed after someone attempted the actual task |
| Handoff | Whether another person can operate and modify the system |

Use synthetic or openly licensed data for a public example. Clearly label simulated customers and results. A polished demonstration with invented impact numbers is weaker than an honest small deployment with documented limitations.

The recommended progression is a vertical slice, then reliability, then broader coverage. Start with the simplest implementation that can meet the acceptance criteria. Expand the architecture only when an observed failure or constraint justifies it. Anthropic's guidance explicitly recognises that agents trade additional cost and latency for potential capability. [Building effective agents](https://www.anthropic.com/engineering/building-effective-agents).

For agent evaluation, distinguish a convincing response from the actual state of the system. Test the result and inspect the trace when it explains a failure. Do not force a single tool sequence if multiple sequences validly complete the task. [Demystifying evals](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents).

## Current technical priorities

Integration work benefits from version awareness. New protocol designs do not instantly remove older clients. Record the SDK version, protocol revision, authentication scheme, and supported peer versions; then test the customer configuration. The MCP release and SDK compatibility notes should be read together. [July 2026 MCP release](https://blog.modelcontextprotocol.io/posts/2026-07-28/), [SDK versioning](https://csharp.sdk.modelcontextprotocol.io/v2/versioning.html).

Retrieval architecture remains a choice. Freshness, permissions, corpus structure, response time, and evaluation results should determine whether indexed retrieval, live exploration, or a hybrid approach is appropriate. The available engineering guidance does not substantiate the previous guide's assertion that indexed search has been broadly displaced. [Context engineering](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents).

Security has to exist outside the model's prose instructions. Limit available operations, enforce access in downstream systems, and exercise cross-user access and unintended side effects. Retrieval and fine-tuning do not remove prompt-injection risk. [OWASP excessive agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/), [OWASP prompt injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/).

## Interviews and compensation

Use the site's employer tabs as preparation guidance, not as a guaranteed sequence of rounds. OpenAI's public guide establishes that assessment formats and AI permissions can vary. Anthropic has explicit candidate guidance. Sierra publishes an engineering process with a system-design screen and an AI-assisted onsite. These are better sources than an unattributed list of supposedly repeated interview questions. [OpenAI interview guide](https://openai.com/interview-guide/), [Anthropic candidate guidance](https://www.anthropic.com/candidate-ai-guidance), [Sierra engineering interview](https://sierra.ai/blog/the-ai-native-interview).

For compensation, the site preserves seven employer ranges with the exact role and pay-composition note. It does not combine base, OTE, and equity into a single market statistic. Ask for the employing entity, location, level, guaranteed cash, variable plan, and equity instrument separately. The earlier use of labour-condition filings as “payroll” was incorrect: they are employer attestations. [US Department of Labor](https://www.dol.gov/agencies/whd/workers/h1b).

## Global opportunity map

Current employer pages show that FDE-family work is distributed across multiple regions, but the market is not globally interchangeable. OpenAI publishes direct FDE roles in US cities and in locations including London, Madrid, Seoul, Singapore, Sydney, and Tokyo. Its Tokyo posting requires Japanese and English and describes travel mainly within Japan; its Sydney posting is hybrid and requires travel up to 50%. These are local roles inside a global organisation, not worldwide-remote offers. [OpenAI Tokyo](https://openai.com/careers/forward-deployed-engineer-tokyo-tokyo-japan/), [OpenAI Sydney](https://openai.com/careers/forward-deployed-engineer-sydney-sydney-australia/).

Europe has direct FDE-family evidence from employers including Palantir and Anthropic. The existing Poland leads remain useful, but several are indexed rather than independently confirmed as open. Country eligibility, language, cross-border travel, and hybrid attendance need to be checked on the exact requisition. [Palantir London](https://jobs.lever.co/palantir/5168e8fd-fec1-4fea-b7a1-81bdaea65850), [Anthropic Munich](https://job-boards.greenhouse.io/anthropic/jobs/5391016008).

India has several current direct-title examples. FourKites lists a senior customer engineer / FDE role in Chennai or remote within India; DevRev lists an FDE role in Bangalore focused on implementation partners; Razorpay lists a Bengaluru FDE role for strategic merchants. These roles share end-to-end delivery but differ in product, domain, customer model, and seniority. [FourKites](https://job-boards.greenhouse.io/fourkites/jobs/7667786), [DevRev](https://job-boards.greenhouse.io/devrev/jobs/6174459004), [Razorpay](https://job-boards.greenhouse.io/razorpaysoftwareprivatelimited/jobs/4723067005).

Latin American evidence includes Telnyx's remote LATAM role focused on Bogotá, Mexico City, and São Paulo, Handoff's São Paulo hybrid role, and Caylent's Mexico posting. The Telnyx role explicitly requires professional Spanish or Portuguese plus English and 10–30% regional travel. Handoff's application identifies contractor status and USD compensation. These details show why “remote” and “paid in USD” do not settle the employment model. [Telnyx](https://job-boards.greenhouse.io/telnyx54/jobs/7776490003), [Handoff](https://jobs.lever.co/handoff/5f222a16-e598-4b0c-b57b-8597e700bc61), [Caylent](https://job-boards.greenhouse.io/caylent/jobs/5976575004).

The Middle East and Africa also have concrete but more role-specific evidence. OpenAI's Abu Dhabi Applied AI Engineer is an adjacent title with end-to-end customer deployment responsibilities. Sand Technologies' Kenya-titled FDE posting describes a long-duration, permanently embedded public-sector health role that states four client-site days weekly and a 12–24+ month commitment. One sentence inside that posting refers to Malawi, so the deployment country needs direct confirmation. These examples should not be generalized into regional norms. [OpenAI Abu Dhabi](https://openai.com/careers/applied-ai-engineer-abu-dhabi-uae/), [Sand Technologies Kenya](https://job-boards.greenhouse.io/sandtechholdingslimited/jobs/4836535101).

A useful global application shortlist records location eligibility, language, travel, seniority, employment model, currency, hybrid or client-site attendance, clearance or sector constraints, and the reason prior work matches the role. Rank on those constraints before brand recognition. Compare packages in their published currency and legal context before attempting an exchange-rate comparison.

## Boundaries of the evidence

The directory contains 57 employers: twelve entries refreshed from employer pages or APIs, four indexed leads, and 41 older snapshots that require a new check. Historical rows are not represented as live vacancies. The source register identifies access limitations. Published company descriptions remain recruiting or marketing material and do not independently verify the employee experience.

The posting checker is an editorial keyword heuristic. Its percentages are score shares, not probabilities. Skill weights, study hours, readiness progress, project criteria, career advice, and daily schedules are planning aids. They do not estimate hiring odds or predict career outcomes.

Another limitation concerns automation: evidence of faster implementation and evidence of continued hiring can both be true. Neither identifies the future number of FDE jobs. The practical response is to develop the ability to specify, validate, integrate, and operate useful systems, while measuring how tools change the work.

## Learning curriculum

The companion [learning path](learning-path.md) adds six stages, 27 milestones, and 25 learning resources. It covers foundations, discovery, integration, AI evaluation, production operations, and applications. Exercises, sequencing, and effort estimates are editorial; resources include official documentation and university course material.

## Benmore and Agent Factory additions

The [source-method guide](source-methods.md) pairs Richard Buehling’s supplied practitioner book with the [Agent Factory orientation](https://agentfactory.panaversity.org/docs/about) and four targeted chapters. These sources strengthen delivery practice, not the job-market estimates above. Nine added milestones cover directing AI, discovery debt, specifications, workflow redesign, governed knowledge, recurring workers, prototype gaps, support ownership, and an engagement proposal.

The [field workbook](field-workbook.md) provides original project templates. Six evidence gates in the interactive case connect those templates to discovery, framing, integration, evaluation, launch, and handoff. Source methods, fictional examples, and customer-story evidence are labeled separately.

## Sources

The complete [source register](sources.json) records 109 sources: 108 web references and one supplied book with publisher, title, publication date where available, original URL, and access notes. Inline links above support the particular findings discussed here. The site's Sources section contains the same register and a correction log.