FDE: The Motion, and the Method Underneath
The pillar this reading feeds stands on one talk, and that talk is about a market — which quadrant needs forward deployed engineers, what they cost, what they earn. It never says what the engineer does on the Tuesday after the contract is signed. The discussion here answers exactly that, and it is the weakest-provenance source this site has read: no named author, a channel compiling four talks it names but does not link, numbers relayed rather than cited. That is why it sits in this room instead of on a sheet. A record whose attribution is its worst property can still be examined — and six of the nine bands below are extended rather than corroborated, because the two sources barely touch. Where they do touch, they disagree once, and the disagreement is the useful part.
Primary source
Forward Deployed Engineering 101Kevin Bai, Anthropic
Recorded talk, 17 min — AI Engineer
Discussion
This Is How Forward Deployed Engineering Is Actually DoneAI LABS
Recorded video, 15 min — a channel compilation, author not identified
The examination — what the discussion did to each claim
- Corroborated
- The discussion agreed with the report.
- Contested
- The discussion pushed back on the report.
- Extended
- The discussion worked out more than the report stated.
- Unresolved
- The report did not say enough to settle it.
A washed band is unsettled — the two sources disagree, or the report does not say enough to decide. Timestamps are positions in the discussion video. It is published by a channel and names no author, so nothing here is attributed to a person. It was read from YouTube's automatic caption track, which mis-transcribes two of the proper nouns it says aloud; quotations elide those words rather than correct them.
Claim 1. The quadrant, and the Tuesday
ExtendedThe record
The whole 101 turns on a 2×2 of what you sell against who buys it. Three quadrants have worn playbooks; the fourth does not. “You only need FTE if you are in this weird unique situation of Palunteer where you are having to sell something very technical to a non-technical buyer.” The talk establishes when a company must staff the function. It never describes a day of the work.
When your buyer isn't technical — 5:03
The testimony
- 1:17
No matter what kind of process it is, there are three things working in it. Either it is a human who's doing the task or it is software that follows hard set rules to do the task or it is an AI.
- 1:28
And an FDE is someone who actually knows which of those three should handle which task in the process.
Reading
Two definitions of the same role, and they do not compete. One is a market condition — a company is in the quadrant or it isn't. The other is a job description narrow enough to do on a Tuesday: look at a process, decide which of three things should run each step. This pillar's six sheets are entirely the first kind. Nothing on this site, on any pillar, says how a step gets assigned to a model in the first place.
Lands onThe 2×2 — the quadrant with no playbook- 1:17
Claim 2. Left to the customer
CorroboratedThe record
The failure the whole motion exists to prevent is stated twice. On the platform: “your success is determined by how well your customers can use your particular piece of software,” so the customer pays once for the software and again in the time it takes to train their people — “that is a terrible way to do business.” And at the end, generalised past Palantir: “if you leave the success or failure of your product to their hands and to their ability to implement, I … assure you this is not … going to be an easy motion.”
What Palantir does, and where FDE fits — 2:36, and What has changed since Palantir — 12:26
The testimony
- 3:48
Last year MIT ran a study on exactly this. They looked at 300 AI projects and surveyed hundreds of people inside the companies that ran them, and they found that 95% of those projects produced no measurable return at all.
- 4:02
And MIT's own conclusion was that the companies themselves caused the failures, not the models.
- 4:29
… he says AI is just getting slapped on top of broken processes because nobody actually looks at the process first.
Reading
This is the only band in the reading where the two sources are saying one thing, and it happens to be the sentence this whole site was built to argue. The primary reaches it from the vendor's side: leave implementation to the buyer and the platform sits on the shelf. The discussion reaches it from a survey of three hundred attempts and lands on the same subject — the companies, not the models. Read the second clause slowly, because it is the thesis line in someone else's words: when the thing fails, the model was not the variable.
Lands onThe Harness — eleven primitives, one chain of limitations- 3:48
Claim 3. The demand, and how well it is evidenced
UnresolvedThe record
The talk offers one measurement of the function's spread, and it is the speaker's own: he joined Rippling as the first person on its forward deployed team, and “we grew it to … around 25 in a year.” That is the entire claim about scale. Every other number in the talk is about contract value, not headcount.
Introduction: what this 101 covers — 0:23
The testimony
- 2:55
Job postings for this role are up 729% in a year. AWS put a billion dollars into building a whole department of forward deployed engineers. And OpenAI's own FDE team went from two people in January to 39.
- 3:11
And all of that happened really fast, because three years ago basically nobody needed an FDE.
Reading
Five numbers, and the difference between them is the whole reason this source is in this room. One is first-hand — a man reporting the size of a team he built. Four are relayed by a channel that names no source for any of them, in a video whose own framing is that the viewer should go get this job. The direction they agree on is almost certainly right; the site's rule does not care. A figure earns a sheet when the talk credited on that sheet states it and the prose names who said it, and “a channel said the postings were up 729%” fails the second half. The primary cannot settle it, and neither can we.
Lands onEverything Is Agentic — the motion pulls to the centre- 2:55
Claim 4. The step nobody wrote down
ExtendedThe record
What the loaned engineer is for is stated as a posture, not a procedure: “you are sending over really smart people who will go and understand the nature of the customer's business.” The image is the fine-dining waiter, “there to cater to your every need,” who “will figure out how to solve you the problem and then build you the software.” How the understanding is actually obtained is left open.
Selling a solution, not a product or a service — 3:00, and The business case — 6:06
The testimony
- 5:05
And Palantir lost a whole year to one of those workarounds. A move onto a new file format was stuck because one engineer kept saying the new format was worse, and nobody could work out why until someone actually watched her work.
- 5:17
Turns out she'd been checking the data by double-clicking the files open, and the new format had nothing you could double-click.
- 9:04
First, you need to watch how a job is really done and write down every single step of it in the order it actually happens.
- 9:12
Then you need to ask why each of those steps is done the way it is. And if nobody can give you a concrete reason, that's usually a work around somebody put in years ago that nobody has questioned since.
- 12:49
And then that document turned out to be most of the build. Those written up processes became the instructions the agents follow and the knowledge the chatbot answers from.
Reading
The waiter metaphor says the engineer finds out what the customer needs. The discussion says how, and the how is a document: every step in the order it happens, then a why against each one, and the steps with no why are the ones somebody bolted on years ago. The last quotation is the part worth keeping, because it is the discussion describing its own build rather than relaying someone else's — the write-up became the instructions. That closes a hole this site has never filled. The harness pillar opens on instructions and treats them as given; here is where they come from, and it is a person sitting next to somebody for an hour.
Lands onInstructions — where the rules actually come from- 5:05
Claim 5. Assembly presumes an allocation
ExtendedThe record
The answer to “you can't maintain a hundred bespoke builds” is that the engineers never start from zero. “They are never writing software from scratch.” There is already a set of primitives “on top of which they could assemble them into some application, some workflow, some solution that is arbitrarily valuable to their customers.” Assembly, not authorship — and the talk stops there, without saying what decides which primitive a given step gets.
FDE as a partnership on a reusable platform — 8:45
The testimony
- 9:36
So you need to run every step you wrote down through three filters. The first is whether that step follows a fixed rule and has to come out right every single time. And if it does, it just stays as ordinary software and doesn't need to be automated with AI.
- 9:48
The second is whether somebody has to read something messy and make a judgment call on it, because that's the part the model is actually for. And the third is what it costs when it goes wrong, because if getting it wrong is expensive, that step stays with a person even when a model could do it.
- 10:06
Some steps stay as software, some go to the model, and some stay with the person.
Reading
A sorting rule with three destinations, and the third filter is one this site already draws from a different direction. On the harness pillar, Ken Ono's roles are laid on an axis of what a wrong answer costs, because verification got expensive exactly where the answer got cheap. Here the same axis is used before the build rather than after it: an expensive mistake keeps the step human even when a model could do it. Put the two together and the site's own question gains a step in front of it. Before asking which harness layer ran out of road, ask whether the step belonged to a harness at all.
Lands onThe Platform of Primitives — assembly, not authorship- 9:36
Claim 6. Centralize, or reach
ContestedThe record
The product the whole motion was built to deliver is a migration. Foundry “enables organizations of arbitrary size to centralize all of their data in one place to create an ontology … to create proper nouns out of their data,” so that instead of “table one, table two, table three,” a company has one source of truth for warehouses, and applications are then built on top of that. The value follows the data into the platform.
What Palantir does, and where FDE fits — 1:47
The testimony
- 6:35
The second step is to build on top of whatever the business already runs, and with AI that matters way more than it used to because an agent is only worth what it can reach.
- 6:42
So, if a team already keeps all their work in Notion, you don't build them a separate system and move all of that across just so an agent can read it more easily. You connect the agent to their Notion with an MCP, and everyone carries on working the way they always did.
- 7:00
One of … clients had spent $5 million and 5 years getting onto their finance system, so moving them off of the system was never going to happen.
Reading
The only place the two sources touch the same decision, and they give opposite instructions. Move the data to the platform, or leave it and give the agent a way to reach it. The opposition is partly generational and worth saying so: a platform sold in 2004 had no cheap way to reach into a customer's systems, so centralising was the only route to a coherent view, and the ontology it produced is the primitive that makes everything else in the primary assemblable. What changed is the price of the integration surface, not the price of the migration. Both sources are answering one question — what does the agent get to touch — and this site puts that question in exactly one layer.
Lands onTool Interfaces — what the agent can reach- 6:35
Claim 7. Trust is the schedule
ExtendedThe record
The offer that closes the deal is framed as a removal of work from the buyer: rather than trusting that they will “spend the time to … not only buy your platform and use it,” you say — “Hey, here's the setup. We will loan you some really good engineers that you don't have to hire, recruit, manage or retain.” The training tax the buyer used to pay is deleted, and the talk treats it as deleted.
When your buyer isn't technical — 5:33
The testimony
- 7:12
The third step is to not change the way people work more than you have to. If somebody's been running an 11-step process for years and you hand them a one-step version, they stop using it. They used to check the work as they went, and now the middle is gone, and they're being asked to trust an answer that just appeared.
- 7:31
You let the agent do the work inside each step, but you keep the steps there so people can still have something they can see, which would let them know the answer is right.
- 7:40
And the last step is to budget way more time for trust than for building. At that same bank … the technical side was finished in 6 to 8 weeks and then it took another 4 months of pilots and testing before the advisors would actually rely on it.
Reading
The tax is not deleted, it is moved onto the vendor's schedule — six to eight weeks of building and four months of earning the right to be relied on. And the mechanism the discussion gives for that is this site's observability argument pointed at a person instead of a system. The harness pillar asks for receipts so the run can be checked; here the eleven steps stay on the screen so the human can still tell the answer is right. Collapse the process to one step and you have not simplified their job, you have removed the only instrument they had. A build that is technically finished and not yet trusted is not finished.
Lands onVerification & Observability — the receipts a person can read- 7:12
Claim 8. What the outcome is worth, in whose ledger
ExtendedThe record
The contract names a business result, not a feature list, because that is the only layer the buyer can evaluate: “if you're working in CPG you care about … getting more placement on the shelves or you care about higher throughput of sales, you don't really care how the data is organized and nor should you care.” The talk establishes what is sold. It does not say who counts it afterwards, or against what.
Selling a solution, not a product or a service — 3:26
The testimony
- 11:21
And last, you have to put a number on what it was actually worth, and you should be able to say whether it brought money in or took a cost out or made a risk smaller, because nothing else counts.
- 11:31
Now, Cursor had somebody complain to them that an agent was costing him $2,000 a day. And when they asked him what that agent was actually doing, turns out it was just picking which engineer to send out to fix broken equipment. So, then they asked him what it was costing him to send the wrong person out, and it was more than 2,000 a day.
Reading
The pillar argues that selling an outcome aligns the contract with the layer the buyer can judge. This adds the part that comes after signing, which is that the outcome has to keep being counted against the right denominator. The complaint was real and the number was real; it was simply measured against zero. What the agent replaced had a price too, and nobody had put it on the same page. Note the shape of the error, because it recurs across this site: the visible cost of the new thing beats the invisible cost of the old one, every time, until somebody writes both down.
Lands onSelling Outcomes — the result, not the licence- 11:21
Claim 9. The profile, and the anti-profile
ExtendedThe record
The hiring bar is given as an intersection and nothing more: “a FTE is nothing more than a customerfacing software engineer … a person who you would hire as a software engineer on your team, but at the same time, you would trust them in front of a customer in some shape or capacity.” Asked directly for the perfect profile, the talk gives the intersection and then runs out of time — “the rest you'll have to figure out as you go.”
Q&A — 16:59
The testimony
- 13:08
One of the companies that hires for exactly this job wants people who are wide across business, process, and technology with one area they're genuinely deep in. And they say the business half gets taught on the job while the technical half doesn't.
- 13:29
… a Palantir exec says that the person who fails at this job is the careful engineer who wants code that still holds up in 10 years. The job is getting something rough in front of a real user quickly…
Reading
An intersection tells you who to consider; an anti-profile tells you who to stop. The second is the more useful instrument and the more poorly evidenced one — an unnamed executive, relayed by a channel, with no way to check the sentence was ever said. Take it as a hypothesis rather than a finding, and note that it is also the sharpest disagreement in the reading that nobody stages: the primary's bar is a software engineer you would put in a boardroom, and this says the software engineer's instinct for durable code is the thing that gets them fired. Both can be true only if the durability belongs to the platform underneath and never to the deployment on top — which is the primary's own sorting rule, arriving from the other end.
Lands onThe Two Questions — need, platform, and the profile- 13:08
Where this reading lands
Nine bands, and only one of them is agreement. That is the shape of a source that continues a talk rather than answering it: the primary explains why a company staffs the function, the discussion explains what the person does once staffed, and the two meet at exactly one decision — whether the agent goes to the customer's data or the data comes to the platform. Read as a career video, this is a channel selling a roadmap and a community, and it should be read that way. Read as a record, its evidence divides cleanly in two, and the site's own rule does the dividing. What it observed — a step nobody documented, a process kept visible so a person can still tell it worked, a write-up that turned out to be most of the build — is first-hand, specific, and belongs to the people it names. What it counted — 729%, a billion dollars, two to thirty-nine — is relayed, uncited, and stays in the unresolved band no matter how plausible it reads. The first kind is why this entry exists. The second is why it is an entry and not a sheet.
06The pillar this reading feedsForward Deployed