What Does a Forward Deployed Engineer Do? Career Guide
What an AI forward deployed engineer (agent operator) does, claimed pay, how to find work agents can automate, and how to start with Claude Code or Codex.
For Engineering and Customer Success

What is a forward deployed engineer (AI agent operator)?
A forward deployed engineer (FDE), in the AI sense used here, is someone who goes inside a business, finds work that AI agents can do, and builds systems so the agents do it. The job is less about writing clever code and more about understanding how a company actually runs. You interview the people who own processes, sort their tasks into what an agent can handle alone and what needs a person working with an agent, then build and tune the setup. Some people also call this role an "agent operator."
TL;DR
- An FDE's core work is finding and automating real business processes, not inventing new AI models.
- The method is to interview process owners, then sort tasks into agent-capable (an agent does 100% of it) and collaborative (a person plus an agent).
- For small and mid-size companies, the creator behind this framework says the easiest wins are in growth work: reach, acquisition, conversion and expansion.
- You can document a process by recording or screenshotting someone's work and feeding the log to a model that turns it into a standard operating procedure (SOP).
- You can start with a coding agent such as Claude Code or Codex, which the source priced at about $20 a month each.
- Pay figures cited for the role ($200,000 to $250,000 median) come from one practitioner, so treat them as a rough signal and check current job listings.
- The hard part is getting into businesses and getting people to explain their work, not the building.
Why is this role getting attention now?
The argument comes from a practitioner who says he has spent seven years implementing AI agents in businesses. It rests on a few ideas worth knowing, even if you don't accept all of them.
First, AI agents can work unsupervised for longer and longer. The research group METR measures this with a "time horizon": how long a task an AI can complete at human-level quality about half the time. The source says METR found that figure doubling roughly every seven months as of March 2025, and that people working on the technology expect the trend to continue.
Second, pay tends to follow the bottleneck. Whatever is scarce in a system gets rewarded. Writing software used to be scarce. The source argues it no longer is, so pay for routine programming has dropped. Doctors and lawyers are paid well because demand for them outstrips supply.
Third, the "centaur" idea: a human supplies judgment while AI does the volume. The source cites Anthropic internally finding that AI agents spend six to seven hours on tasks for every hour humans do.
If those ideas hold, the scarce person is the one who can connect agents to real business problems. That is the FDE.
What does a forward deployed engineer do all day?
According to the source, most of the day is talking to people. The building is delegated to coding agents, and the operator frames requests with enough context for the agent to build what the business needs.
Examples the source gives of things an FDE might build:
- A Python system that queues and triages a client's email every morning, connected to their Gmail.
- A system that uses AI to generate assets or proposals for a business that is held back by producing them (a home services company is the example).
A typical engagement looks like this:
- Meet the business owner (for a small company) or department heads (for a bigger one).
- Have them walk through their processes.
- Sort each task into one of two buckets.
- Automate the first bucket and train people on the second.
- Tune the outputs until they match what the business needs.
How do you tell which tasks agents can do?
The split is simple, and it is the center of the job.
| Bucket | What it means | What you do | Expected effect (per the source) |
|---|---|---|---|
| Agent-capable | An agent can do 100% of the task without a human | Hand it to agents completely | Uplift of "several thousand percentage points" |
| Collaborative | The task needs a human working with an agent | Train the person to use agents as a thought partner or sounding board | Improvement of a few hundred percentage points |
Those percentages are the creator's claims, not measured results, so read them as an upper-end sales pitch. The practical point stands: some work can be fully handed off, and some work gets better when a person is given better tools. Treating both the same way is how projects fail.
Be honest with clients about the first bucket. Fully automating a task can reduce the labor a company needs. The source says as much. For the second bucket, the source suggests the freed-up person can move into more valuable work.
Where should an FDE look first?
The source says most small and mid-size companies (he suggests under about 100 staff) struggle more on the front end of the business than the back end. So he focuses on growth, using the acronym RACE:
- Reach: getting in front of potential customers.
- Acquisition: turning attention into leads or sign-ups.
- Conversion: turning leads into paying customers.
- Expansion: selling more to existing customers.
His claim is that RACE work is the easiest and most economical place to start. It is also easy to explain to an owner: more leads, faster follow-up, fewer dropped proposals. A good first question for any owner is "Where does a lead or customer wait on a person?"
How do you map a process so an agent can run it?
You need every step the person takes, including the small ones they've stopped noticing. The source's method:
- Ask permission. Only record someone's work if they agree and the company allows it.
- Capture the work. Shadow the person and record it with whatever they are comfortable with. If the work is entirely on a computer, screenshot it every few seconds.
- Build a log. Feed the recording or screenshots to a model such as Claude.
- Turn the log into an SOP. Ask the model to write a step-by-step procedure for each task.
- Let an agent run it. The source mentions browser use (Claude working in a web browser) and computer use (Claude controlling the cursor).
- Tune. It won't be perfect on the first try. Work with the process owner until the output matches what they need.
For meetings with owners, the source also suggests an auto-transcribing device (he names Pocket, and says he isn't sponsored) so the conversation becomes a transcript you can hand to a coding agent right away.
How do you get started as a forward deployed engineer?
Start by using a coding agent. The source says Claude Code and Codex each cost about $20 a month at the time of the video. Both can run in a terminal, and Claude Code also has a desktop version. You can talk to it by voice instead of typing. He also runs a tool called Orca to run several agents side by side, but says you don't need it.
A practical path:
- Pick one real process. Use your own job or a friend's small business. Email triage, lead follow-up and proposal drafting are all in the source's territory.
- Document it first. Write the SOP, or record the work and have a model write it.
- Build one small automation with a coding agent and measure the time saved.
- Write it up as a short case study: before, after, what the person does now.
- Find your next business. The source says the bottleneck is not building. It is getting into businesses and getting owners to explain their processes.
Your own network is the cheapest source of first clients. Ask people you already know who run or manage a small team.
How much does a forward deployed engineer make?
The source says the median FDE in the United States makes about $200,000 to $250,000 a year in total pay, with the top 10% close to $400,000. He also recalls Anthropic posting a role at around $300,000 to $350,000. He says Anthropic announced a plan to train 10,000 such engineers by the end of 2027 and called them the hardest talent to find.
Treat these as the creator's numbers. They come from one person who also runs a business in this space and sells the idea. Check current listings for "forward deployed engineer" at AI labs and software companies, and notice that those roles often expect real engineering skills. A freelance or consulting version of the job, serving small businesses, will pay differently and less predictably.
Is this a safe career bet?
It is a reasonable bet with real caveats.
- The forecast is a forecast. The METR trend and the claim that most salaries will fall are projections, and the source's confidence is high.
- The role itself may change. If agents get better at finding processes, the FDE's job shifts too.
- Trust is the product. You will see how a company works, handle client data and sometimes automate coworkers' tasks. Consent, privacy and clear communication matter as much as technical skill.
- Domain knowledge helps. An accountant who can automate month-end close, or a recruiter who can automate screening, has a head start over a generalist.
What does this mean for you?
If you work in an operations, marketing, finance or sales role, you already know where the repetitive work lives. That knowledge is the scarce part. Start by sorting your own tasks into agent-capable and collaborative, and try automating one.
If you manage a team, expect someone (an employee or a consultant) to ask you to explain your processes. The clearer your SOPs, the easier it is to decide what to automate. If you want to move up from there, see the path from director to C-suite and what AI changes along the way.
If you are a developer, the shift is from writing code to scoping problems and talking to owners.
What could you build to run an FDE-style project?
An FDE ends up juggling interview notes, process lists and task statuses, usually in a spreadsheet. With Remy, you can describe a small process-intake app in plain English and have it built and deployed as a full-stack app with a login. For example: a form where process owners list their steps, a field that tags each task as agent-capable or collaborative, a RACE category, and a dashboard showing which processes are mapped, automated or in review. Remy writes a spec, builds the app and deploys it to a live URL, so you can share it with a client instead of sending another spreadsheet.
Frequently Asked Questions
Do I need to be a software engineer to become a forward deployed engineer?
The source suggests you don't need to write code by hand, since coding agents like Claude Code and Codex do the building. You do need to understand business processes, explain them clearly and check that what the agent built works. Roles at AI labs may still ask for engineering experience.
What does RACE stand for?
RACE stands for reach, acquisition, conversion and expansion. These are the growth-side parts of a business. The source's advice is that small and mid-size companies usually struggle there more than in back-office work, so it is a good place to look for automation first.
Is it okay to record an employee's work?
Only with their agreement and within your employer's or client's policies. The source says to record only if the person is comfortable with it, and to use a method they accept. For digital work, periodic screenshots can be enough.
What tools do I need to start?
A coding agent is enough. The source names Claude Code and Codex, each about $20 a month at the time. Extra tools, such as a way to run several agents at once or a transcription device for meetings, are optional.
How we source, write, and correct our articles. Editorial standards
Related articles

Can a Coding Agent Run a Robot? Stanford's VIA Experiment

Gemini Carbon and Hydrogen: Does the Orchestrator Save Quota?

What Are AI Software Factories? The 5 Types Compared
