This was the week "forward-deployed engineer" stopped being a job title and turned into a Rorschach test. The same two words, used to mean at least three different things, with a pile of likes behind each version. I have spent years doing the work the phrase describes, and watching the timeline fight about it, I kept landing on the same thought: everyone is arguing about the label and nobody is describing the job.
I run Black Matter VC. I build and operate AI systems inside venture funds — Systemiq, Northzone, Clean Energy Ventures, Active Impact, FirstMinute. Not slideware. Production systems partners actually use on a Monday and complain about on a Tuesday. Before that I ran no-code engineering at Calm Company Fund for four years. So when my whole feed started talking about embedding engineers inside companies to make AI land, my reaction was not "interesting trend." It was: that is the job I already have, and the framing around it is off.
IThe week the words took over
Here is the sequence, because the order tells you something.
It opened with Aaron Levie. Box's CEO called forward-deployed engineering the most-coveted role in tech, compared the opening to the cloud-migration wave, and pointed career counselors at CS students with operational instinct. I clocked it at 646 likes in my Monday digest and flagged it as the post that set the week's tone.
Then came the counter-signal. Ethan Mollick, the Wharton professor who writes Co-Intelligence, posted the line that traveled furthest of any of them: you will know the labs believe in ASI the day they disband their newly formed forward-deployed engineering groups. I had it at 885 likes by Monday. Same vocabulary as Levie, opposite conclusion. The logic is clean and a little brutal. If superhuman models are months away, why are you frantically hiring humans to sit next to the deployment?
Two posts, one phrase, pointing in opposite directions. That gap is the whole essay.
IIThree reads, all of them half-right
By midweek the timeline had sorted into three camps. Each is worth naming, because each is half-right, and the half it misses is the interesting part.
The first read is Levie's: forward-deployed engineering is the career of the decade, a goldrush role for people who can sit between a model and a messy org. He is right that the demand is real. He misses that "most-coveted role" frames it as a job you take at someone else's company, when the actual money is in running it as your own practice.
The second read came from Harry Stebbings, who runs 20VC. He called the labs deploying engineers into enterprises "an obvious and right move" and read it as a flashing sign of the vertical-AI opportunity: each deployment surfaces a gap a purpose-built product eventually fills. 162 likes, 53,000 views. I agree with him almost all the way. I said nearly the same thing in my own post this week: the consulting leg is field research, not a business model. Where I would push back is the next step. He treats the services layer as scaffolding you discard once the product exists. Some of it is. A lot of it is load-bearing and never comes down.
The third read is Mollick's disband-at-ASI line, and he extended it the same week with a second post I think matters more than the first: insourcing. His argument is that companies will stop paying outside vendors for legal, marketing, and software, and pull the work in-house, because a smaller internal team can capture those gains directly with AI. 381 likes and 57 people arguing in the replies. This is the strongest version of the bear case against people like me. If AI makes internal teams that productive, the forward-deployed outsider is a temporary patch.
Three labels: career goldrush, vertical-AI signpost, temporary patch. One shared assumption underneath all of them. Each treats the forward-deployed engineer as a phase, a role that exists only because the technology is mid-transition and will resolve once it matures. I do not think that is what is happening. I think the role exists because of something that does not resolve.
IIIWhat the labs are actually telling you
If you want to know where the value is settling, do not read the takes. Watch what the labs spend money on.
This week Anthropic acquired Stainless, the SDK and MCP-server platform that has quietly powered every Anthropic SDK since their earliest API days. The post did 891,000 views and roughly 2,700 likes. It tripped the anomaly flag in our own tracking, which is usually a sign the audience knows something matters before the commentary catches up.
Think about that for a second. The most capable model lab in the world just bought a company whose entire job is making the model easy to wire into someone else's stack. Not a research team. Not compute. The integration layer — the plumbing between the model and the place the model has to actually work.
That is the tell. Labs are hiring forward-deployed engineers and buying SDK platforms not because the models are weak, but because the model was never the hard part. Wiring intelligence into one organization's data, permissions, edge cases, and people is the hard part. It is also the part that pays.
IVWhat it looks like from the fund-platform side
Let me give you the version I actually live, because the abstract debate has a concrete shape.
Every enterprise AI deployment I have shipped behaves the same way after launch: the clients keep calling. Not because the build was broken. Because the model drifts, the context shifts, an edge case nobody scoped shows up in week three. You did not hand off a piece of software. You stood up a system that has to be operated. The part the "it's a stopgap" camp misses is that the ongoing loop is not the cost of the work. It is where the value gets created.
Balaji Srinivasan put the principle in one line this week, and it earned 845 likes for good reason: every AI agent ultimately has a human principal. An agent running inside a fund's deal pipeline is not autonomous. It acts for a partner who has to trust its output enough to move on it. Someone has to own that trust, tune it, and answer for it when it slips. That someone is the forward-deployed engineer. A better model does not remove the role. It makes more decisions, which raises the stakes on the loop instead of lowering them.
I watch this land in venture funds before it lands anywhere else. AI shows up in the platform team, the people who run deal flow, CRM, and LP reporting, long before it shows up in IT. A fund does not want a seat license. It wants the deal-memo screener that still works after the partners change their thesis, the LP-reporting pipeline that survives a new fund structure, the brief that is still right when the portfolio doubles. That is not a product you install. It is a system you run.
VWhat this means Monday morning
If you are a builder reading this as a career question, Levie is directionally right but the framing is small. The high-leverage move is not taking the most-coveted role at a lab or an enterprise. It is owning the operating loop for a vertical you understand cold — install the system, then stay on to run it. That is a practice, not a job.
If you run a fund or a platform team, the lesson is sharper. The next time a vendor sells you an AI "solution" you can buy and forget, assume the opposite. The thing you are buying needs an owner. Either you grow one internally, which is Mollick's insourcing path and which I take seriously, or you bring in someone who builds it and operates it without selling you forty seats you will use for a week. Both work. What does not work is treating the deployment as finished the day it ships.
VIThe honest part
Where do the skeptics have a point? Mollick's insourcing argument is the one that keeps me honest. If models keep improving at the rate they have been, a sharp three-person internal team really can absorb work that used to need vendors. I have watched a couple of funds start down that road, and for some of them it is the right call. I say so on the call, even when it costs me the engagement.
What I do not buy is that better models erase the operating layer. Every capability jump I have shipped through did the same thing: it moved the bottleneck, it did not remove it. We went from "can the model do it" to "can we trust what it did," and trust at scale is an operations problem, not a model problem. I flagged the same shift in my digest this week — as generation gets cheap, reviewing what got generated becomes the expensive part. Disband the forward-deployed engineers and you do not get autonomy. You get a stack of confident output nobody owns.
So no, I do not think the labs disband these teams when they believe in ASI. I think they staff them harder, because someone has to stand between a very capable system and a customer who has to live with its mistakes. That is not a phase. That is the job.
VIIWant this for your fund?
If your platform team is staring at an AI rollout that keeps needing attention after launch (the screener that drifts, the reporting pipeline that breaks on a new fund structure), that is the operating loop, and it is exactly what Black Matter builds and runs. I install the system and stay on to operate it: deal flow, CRM, LP reporting, portfolio data, the weekly review cadence. Email michael@blackmatter.vc. $10k/mo flat retainer, no lock-in.
VIIIRead more
I publish a build essay every Saturday and a digest of what actually shifted every Monday at blackmatter.vc/lab — what shipped in agent infra, what flipped, the handful of posts worth your time. If the forward-deployed question is live for you, the digest is the easy follow-up.
I am still arguing this one out in real time, including with myself. If you are running deployments that will not stay shipped, I want to compare notes — what is drifting on you right now?
— Drawing on this week's signal: @levie and @HarryStebbings on the opportunity, @emollick on insourcing and the disband-at-ASI counter, @AnthropicAI on the Stainless acquisition, @balajis on the human principal, plus my own posts of May 18–20.
— Michael Rouveure · 24 MAY 2026
