Your role, end to end
Forward-deployed engineer
From a customer's pain to a system they run after you leave.
On this page
Who this page is for, and how to use it
Engineers who build inside a customer's organisation, or inside another team in their own company, and anyone who hires, manages or works beside one.
Run an engagement end to end, with a template, prompts and the words to say at every step.
- Read the picture: it is the whole job.
- Read the altitude table before your next proof of concept.
- Open the stage you are in.
Frame, deliver, evolve: the same four questions, three times
Rows are an engagement's three stages, in order. Columns are the manual's four phases, P0 to P3.
- FFramethe engagementDays. The client's people watch.
- P01QualifyIs this engagement worth taking?Discovery brief
- P12ScopeWhat will we prove, and who signs?Statement of work
- P23ProveDoes it work on their data?Proof-of-concept sheet
- P34DecideGo, at what cost, on what conditions?Readout and decision
Signed the go decision, by the client's sponsor
- DDeliverthe systemWeeks. The client's people build with you.
- P05MobiliseIs the first slice still right, on real access?Access list and baseline
- P16SignWhat exactly, under whose authority?Signed spec and limits
- P27BuildDoes it meet their bar, slice by slice?Evidence pack
- P38Hand overDoes it run without you?Handover pack
Signed the handover, by the person who will run it
- EEvolvethe relationshipMonths. The client's people run it.
- P09ReframeWhat should come next, here?Next-frame brief
- P110CodifyWhat comes back to the product?Pattern entry
- P211ReuseDoes the reusable version work next time?Asset proof sheet
- P312ReviewIs the relationship healthy, and what did we learn?Quarterly value review
Signed the next frame, or a clean close, by the client's sponsor
Every box opens its step: template, prompts, a worked example. After the review, Frame starts again.
What does a forward-deployed engineer do?
An engineer who works inside a customer's organisation and owns the outcome there, then carries what they learnt back home. The job runs in three stages that are its own initials: Frame, Deliver, Evolve. In the words of three companies that run FDE teams:
You will own discovery, technical scoping, system design, build, and production rollout
OpenAI, Forward Deployed Engineer (FDE), San Francisco · jobs.ashbyhq.com ·Forward Deployed AI Engineers' responsibilities look similar to those of a hands-on AI startup CTO
Palantir, Forward Deployed AI Engineer (Delta), New York · jobs.lever.co ·Identify and codify repeatable deployment patterns and contribute insights back to our Product and Engineering teams
Anthropic, Forward Deployed Engineer, London · job-boards.greenhouse.io ·
You are judged on five things
- Impact: value against the customer's goals, and the system becomes part of their work.
- Delivery: milestones hit, little reopened.
- Reuse: patterns of yours that other deployments use.
- Judgement under pressure.
- Product: field signal that changes a roadmap.
OpenAI's measures for a Deployment Lead in its FDE team1.
One person, six hats, and the decisions stay theirs.
Some weeks you are the whole team. Each hat has a role page here: borrow its steps. What the last column names, you draft and they sign.
| The hat | What you do in it | Borrow these steps | What stays theirs |
|---|---|---|---|
| Product manager | Measure the pain, size the value, cut the first slice | PM steps 1 to 3 | Which pain is worth paying for |
| Solution architect | Map the steps, draft the authority budget | Architect steps 3 and 6 | Every limit, and who approves above it |
| Engineer | Build in their repository, with their engineers | Engineering steps 1 to 5 | Merge rights, and who owns the code |
| QA lead | Build the golden set, the harness and the shadow | QA steps 2, 4 and 7 | What counts as a right answer |
| Platform | Access, environments, the rollback | DevOps steps 2, 3 and 8 | Who runs it after you leave |
| Consultant | Discovery, the statement of work, the readout, the review | This guide, steps 1, 2, 4 and 12 | The decision to buy, and to continue |
At a larger vendor the work is split in two. Palantir pairs its engineers with Deployment Strategists2, and Databricks does the same3. Anthropic adds a Technical Deployment Lead4, and Ramp an AI Solutions Strategist5. Glean puts its founding FDEs in a pod with Forward Deployed PMs
6. Learn both halves: when the pair is one short, you are both.
Wear the consultant's hat as a craft. AWS sets its FDEs against consulting that assesses, recommends, and treats each deployment as a standalone project
7: you build, and you stay until it runs.
Think at the altitude the step needs.
Most FDE mistakes are a right answer at the wrong altitude: a proof built like production, or production built like a proof.
| POCstep 3 · days | MVPsteps 5 to 7, one slice · weeks | Buildstep 7, every slice · weeks | Deploystep 8 · until handover | |
|---|---|---|---|---|
| The question | Can it work at all, on their data? | Will one group of real users get value from the thinnest version? | Will it hold on every slice, at full volume, when things go wrong? | Will it keep working without you? |
| You think about | One hypothesis | One workflow and one slice | The system: every slice, every failure, the cost | The organisation: owners, operations, the contract |
| Code | Throwaway, in a sandbox | Their repository, their identity, the floor tested | Caps in tool signatures, the harness blocking the merge, traces | Runbooks, timed switches, alerts with owners |
| Evidence | A number with its sample size, and what it does not prove | A pass mark for the slice, and agreement in shadow | The lower bound above the pass mark, per slice, and cost per case | The saving and the spend on one line, and drift watched |
| You may skip | Hardening, scale, the interface | Other slices, and any autonomy above drafting | Nothing in scope | Nothing |
| Never skip | Writing down what it does not prove | Their risk owner's signature | A harness the merge cannot pass | The person on their side who will run it |
| Your AI tools | A coding agent writes most of it. You read the evaluation, line by line | The agent builds from a story file. You review every change that touches their systems | The agent writes and the harness proves. Two people read anything that moves money | The agent drafts the runbooks. People rehearse them with a stopwatch |
| Ends with | Go, go with conditions, change course, or stop | A shadow report | The pass mark met, slice by slice | A handover signed by their operator |
| The trap | The demo becomes the success criterion | The first slice becomes the product without a pass mark | Polishing slices nobody uses | Leaving with nobody owning it |
The proof column follows GOV.UK's service manual on alpha: build things just complex enough to let you test different ideas, not production quality code
8, and expect to throw away any code
8. Anthropic's own advice on agents points the same way: finding the simplest solution possible, and only increasing complexity when needed
9.
Your client may sit in your own company.
The twelve steps do not change. Five things do.
| External client | Internal client | |
|---|---|---|
| The agreement | A statement of work, signed by both companies | A charter, signed by the business unit's head |
| Saying no | A change request, sized in days | A trade-off put to the sponsor you share |
| Where learning goes | Your product team, or your practice's library | Your platform's backlog |
| How it ends | Handover, then the contract closes | Handover on the charter's date, or you become their team for good |
| What failure costs | The renewal | Their trust in the next team you send |
Internal clients are real. Palantir's Deployment Strategists may also be deployed to Palantir internal teams and problems
2, and Databricks' AI FDEs support internal subject matter expert (SME) teams
10.
Two crafts, used every week.
The consultancy craft wins and keeps the work. The technical craft makes it run. Each move lives in a step.
The consultancy craft
- Discovery in their data, not the pitch step 1
- A map of five people: sponsor, risk owner, operator, expert, sceptic step 1
- A statement of work a sceptic would sign step 2
- Saying no, and not yet with the condition that changes it step 2
- A readout that leads with what failed step 4
- Change control in one paragraph step 2
- Their names on their decisions step 6
- A value review, the saving beside the spend step 12
The technical craft
- Every access request on day one, with a one-call test step 5
- A walking skeleton through their real system step 7
- Their systems as tools, each with the least privilege step 7
- The floor before the prompt, a checker after risky calls step 7
- Caps in the tool signature, never only in a prompt step 6
- Their golden set, the lower bound, a harness that blocks step 7
- A shadow run beside their staff step 7
- Switches timed before cut-over, and an owner on their side step 8
Your AI tools behave differently in their building.
You work in their repository, under their policy, often on their cloud account. Six dated facts, then five rules.
Six dated facts
- If a repository has AGENTS.md and no CLAUDE.md or CLAUDE.local.md, Claude Code reads AGENTS.md. Codex reads AGENTS.md files before it does any work, from the repository's root down to the folder it starts in, and a nearer file overrides the ones above it.code.claude.com · learn.chatgpt.com
- With an API key or a cloud provider such as Amazon Bedrock you lose cloud sessions, routines, Remote Control and the Chrome integration, and on Bedrock the built-in web search too.code.claude.com
- Computer and browser use, GitLab and self-hosted GitHub Enterprise Server are not yet supported in Codex Cloud environments.learn.chatgpt.com
- A custom connector is called from Anthropic's cloud, so its server must be reachable from the public internet, and one behind a VPN will not connect.support.claude.com
- OpenAI warns against setting
OPENAI_API_KEYorCODEX_API_KEYfor a whole CI job that runs the repository's code, since its build scripts, tests and dependency hooks can read it.learn.chatgpt.com - Without
--bare, a-prun executes the repository's hooks and connects its MCP servers with no trust prompt.code.claude.com
Five rules
- One context file in their repository, AGENTS.md, so both tools read the same rules.
- Only the tools their policy approves, on the account their data may live in.
- No key in a CI job that runs their code.
- Headless runs bare, in any repository you did not write.
- Every document they hand you is untrusted input to your agent.
The Tool guides, with every fact dated →
Ramp's FDE team says it is very heavy users of Cursor and Claude Code
11, and AWS asks its senior FDEs for experience leading agentic and spec-driven development at production scale
12. The tools are part of the job description, which is why the guide gives them a section.
You can start from five places.
Companies hire forward-deployed engineers from university, from an early-career rotation, from engineering, from a systems integrator and from delivery leadership. The bar rises with what you own.
| Start from | Who hires there | What they ask, in their words |
|---|---|---|
| University | Palantir, Forward Deployed Software Engineer, New Grad | Must be graduating in December 2026 or Spring 202713 |
| Early career | Anthropic, a six-month Applied AI rotation | 0-2 years of software engineering, technical consulting, or solutions architecture experience14 |
| Engineering | OpenAI, Forward Deployed Engineer | 5+ years of engineering or technical deployment experience that includes customer-facing work15 |
| A systems integrator | DXC, certified through Anthropic Academy | DXC will recruit engineers from its existing development teams and certify them16 |
| Delivery leadership | OpenAI, Deployment Lead | 7+ years of customer-facing technical delivery leadership1 |
Four rungs
- Builder. Ships one slice in someone else's stack (steps 5 to 7).
- Deployer. Owns one engagement from the statement of work to the handover (steps 1 to 8).
- Lead. Runs several engagements, and owns the value and the relationship (all twelve, across accounts).
- Practice lead. Turns patterns into product and playbooks, and hires (steps 10 to 12, for the practice).
The top two rungs are in the managers' postings. OpenAI's FDE manager must codify what works into tools, playbooks, and roadmap inputs
17. Anthropic's must qualify engagements, scope work, and inform statements of work
18, and build repeatable playbooks, starter repositories, integration templates
18.
Ninety days to your first engagement
- Days 1 to 30, build. A rebooking agent over a public flight dataset, with one tool as an MCP server, a refund cap in the tool's signature with its test, 200 labelled cases and a lower bound.
- Days 31 to 60, frame. Write its discovery brief, statement of work and proof sheet as if for a real sponsor, and have a stranger red-team the statement.
- Days 61 to 90, deliver and evolve. Shadow it against your own labels, time the rollback, and write the handover pack and one pattern entry. Then play the simulator from Day 1 as the whole team: you make all thirteen decisions.
Ten interview questions, by stage →
Ramp finds drive and work ethic the single best predictor of real-world performance
11, and of the 16 FDEs on the team, 7 of us are previous founders
11. Scale wants people who can tell the difference between a customer's stated request, their actual problem, and what the platform should do
19.
Read next.
- What is an FDE? The role, in the words of the companies that hire oneBeginner
- The Deliver stage in depth, with AI-DLC and AIDDIntermediate
- Ten FDE interview questions, by stageAdvanced
Every lesson in the tutorial ends with a row for you: what a forward-deployed engineer does with it.
Play the simulator from Day 1. You make all thirteen decisions as the whole team: an FDE's ninety days in fifteen minutes.
Where every claim on this page comes from
Every claim about the profession is a posting, a page or an article, with the date it was checked. A date turns amber once it is more than sixty days old.
- OpenAI, Deployment Lead (DL), FDE, Tokyo · jobs.ashbyhq.com ·
- Palantir, Deployment Strategist (Echo), New York · jobs.lever.co ·
- Databricks, Deployment Strategist · databricks.com ·
- Anthropic, Technical Deployment Lead · job-boards.greenhouse.io ·
- Ramp, Software Engineer, Forward Deployed AI Solutions · jobs.ashbyhq.com ·
- Glean, Founding Forward Deployed Engineer, New York · job-boards.greenhouse.io ·
- AWS, AWS invests $1 billion to embed AI forward deployed engineers with customers · aboutamazon.com ·
- GOV.UK, Service Manual: how the alpha phase works (updated 8 May 2019) · gov.uk ·
- Anthropic, Building effective agents · anthropic.com ·
- Databricks, AI Engineer, Forward Deployed Engineering (AI FDE) · databricks.com ·
- Ramp, Forward Deployed Engineering · engineering.ramp.com ·
- AWS, Sr Forward Deployed Engineer, AWS Forward Deployed Engineering · amazon.jobs ·
- Palantir, Forward Deployed Software Engineer, New Grad, Commercial (Delta), New York · jobs.lever.co ·
- Anthropic, Associate Applied AI, Rotational Program, London · job-boards.greenhouse.io ·
- OpenAI, Forward Deployed Engineer (FDE), San Francisco · jobs.ashbyhq.com ·
- Anthropic, DXC will integrate Claude into the systems banks, airlines, and other regulated industries rely on · anthropic.com ·
- OpenAI, Manager, Forward Deployed Engineering, New York · jobs.ashbyhq.com ·
- Anthropic, Manager, Forward Deployed Engineering, New York · job-boards.greenhouse.io ·
- Scale AI, Forward Deployed Product Manager, Enterprise · job-boards.greenhouse.io ·
Start where every engagement starts.