The first consensus of the Agent era has emerged: the model is no longer the bottleneck; people are.
In just one week, OpenAI poured in $4 billion, Anthropic embedded itself into FIS headquarters, and Google announced the hiring of hundreds—three AI giants are simultaneously betting on the same role: FDE.
On May 11, 2026, OpenAI announced the establishment of the "OpenAI Deployment Company" with an initial investment of $4 billion. Its core business is simple: "dispatching" engineers into client companies to help them get AI up and running.
Just a week prior, Anthropic embedded its engineering team into fintech giant FIS, aiming to compress anti-money laundering investigations at BMO and Amalgamated Bank "from hours to minutes" by the second half of 2026. A week before that, Google Cloud CEO Thomas Kurian personally took to LinkedIn to recruit "hundreds" of people, a post that garnered 1.3 million views on X.
The role being targeted by all three companies is the same: Forward Deployed Engineer (FDE).
A role that had only been popular for twenty years at the "alternative" software company Palantir has suddenly become the hottest position in the AI industry in 2026. Some are shouting the slogan: FDE is the PMF paradigm of the Agent era.
Is this judgment a profound insight or wishful thinking? To answer this, we must clarify: what exactly is an FDE, why did it suddenly become a necessity in 2026, what is its relationship with "PMF"—and what are its limitations?
I. What is FDE: Not a Sales Engineer, Nor a Consultant
First, let's translate two terms, which are the prerequisites for all subsequent discussions.
PMF (Product-Market Fit) is the "Holy Grail" of Silicon Valley startups. It means your product perfectly meets a real market demand, and the market is willing to pay for it, repurchase it, and spread it by word of mouth. Before finding PMF, a startup feels like it's swimming against the current; after finding it, it's like drifting downstream.
FDE stands for "Forward Deployed Engineer," but it wasn't originally defined that way. The role was invented by Palantir in the early 2000s when their clients were US intelligence agencies—people who "couldn't clearly state what they wanted, wouldn't let you see their data, and whose workflows were constantly changing."
Palantir co-founder Shyam Sankar has a classic quote: "If a problem could be solved by a requirements document, it would have been solved long ago."
So Palantir did something unconventional: they stopped asking clients "what do you want" and instead sent engineers directly into client offices, military bases, and even aircraft assembly shops to write code alongside them. These people were called "Deltas" within Palantir—they had to pass the same engineering interviews but worked in air force bases, bank back offices, and hospital IT systems instead of open offices in Palo Alto.
This differs from three common roles in traditional software companies:
- Sales Engineers handle pre-sales demos and leave once the contract is signed;
- Solutions Architects primarily provide technical consulting and don't write production code;
- Consultants usually provide methodologies and delivery but don't participate in the vendor's product iteration.
The uniqueness of an FDE is that they write the client's production code while feeding back common problems found in client scenarios to the vendor's core product. Palantir's internal description is: "The scope of an FDE's work is like a startup CTO—responsible for a high-risk project end-to-end in a small team."
By 2016, the number of FDEs at Palantir once exceeded that of regular software engineers. The company's entire product form—the Foundry platform—was essentially "distilled" from countless FDE on-site projects. An engineer who served as an FDE at Palantir for seven years summarized this model as "gravel road to paved highway": FDEs build numerous gravel roads at client sites, and the product team identifies the most traveled ones to pave into highways, turning them into platform capabilities.
II. The 2026 Turning Point: Why Three AI Giants Are Betting on FDE Simultaneously
For nearly twenty years, Palantir's model was viewed as an "outlier" in mainstream Silicon Valley—most SaaS companies advised "don't learn from Palantir, the gross margins won't hold up." But in 2026, things suddenly changed.
On May 4, Anthropic and FIS announced a partnership where Anthropic's Applied AI team and FDEs would "embed" within FIS to co-design financial crime AI Agents.
On May 11, OpenAI officially announced the OpenAI Deployment Company (internally codenamed "DeployCo"), with an initial $4 billion investment led by TPG, with participation from 19 investment and consulting firms. Simultaneously, they announced the acquisition of Tomoro, an applied AI consultancy, bringing in about 150 FDEs and deployment experts.
On May 12, Google Cloud CEO Thomas Kurian announced a new "AI-focused organization" within Google Cloud to hire "hundreds" of FDEs. At the time, Google Cloud had 59 related job openings.
Why now? Why all at once? The judgment of the three companies points to one fact: the bottleneck of the Agent era is not the model itself, but the deployment.
Accenture's "Pulse of Change" survey shows that only 32% of business leaders report seeing "sustained, enterprise-wide AI impact." The remaining 68% are in a state of having pilots, PPTs, and demos, but no large-scale delivery. In an IBM survey of 2,000 executives in early 2026, "execution speed" was listed as the third-highest strategic priority.
OpenAI's announcement put this logic bluntly: "Over the past few years, more than a million enterprises have adopted our products and APIs. A pattern is becoming increasingly clear—the winner of the next phase of enterprise AI depends on how effectively a company can deploy this technology into real business scenarios."
There's another set of data worth noting. OpenAI reportedly missed its internal revenue and weekly active user targets in early 2026, while Anthropic and Google Gemini continued to eat into the enterprise market share. OpenAI's Applied Business CEO Fidji Simo called Anthropic's progress a "wake-up call" and said the company must "deliver on productivity scenarios."
In other words, the marginal utility of the AI model's "product power" is declining, but the marginal utility of the engineering capability to "turn models into usable systems" is skyrocketing. No matter how strong the model is, if it can't run within a bank's compliance process, an insurance company's claims system, or a manufacturing MES system, it's just a demo, not a business.
FDE is precisely that converter.
III. Why the Agent Era Has a "Structural Demand" for FDEs
To understand why "Agent" and "FDE" are a perfect match, we need to clarify the fundamental difference between Agents and previous AI forms.
Traditional SaaS products are essentially "tools": you buy Salesforce to get a set of configured sales process templates for your people to use. The boundaries of a tool are clear—what it does and doesn't do is explicitly written in the product manual.
An Agent is about "acting on behalf": you no longer use it; you let it do things for you. An anti-money laundering Agent doesn't just give investigators a better query interface; it helps them complete the entire workflow of "pulling evidence from core systems, cross-referencing known laundering patterns, judging risk levels, and drafting Suspicious Activity Reports (SAR)."
This difference has three consequences:
First, Agents must be deeply embedded in the client's real workflow. To "act on behalf," an Agent must know where the bank's compliance boundaries are, which decisions cannot be automated, how SAR reports should be written to be accepted by regulators, and where internal data is stored. These things aren't in product documents; they are in the client's "institutional muscle memory."
Second, an Agent's failure is a "business failure," not a "functional failure." If a SaaS button is missing, users complain. If an Agent misses a suspicious transaction, the bank gets fined by regulators. This means Agent deployment relies more heavily on "domain knowledge" and "operational context" than any previous generation of software.
Third, the Agent market is one where "there are no mature products to benchmark against, and clients themselves don't know what they want." This is exactly the situation Palantir faced with intelligence agencies. Clients can say "I want AML investigations to be faster," but they can't define "fast," which data sources to use, which steps to automate, or which human decision points to keep. This kind of problem can't be solved with a requirements document; it requires engineers to go in, observe, test, modify, and observe again.
Anthropic's FDE job description clearly outlines this logic: "Build production applications within client systems, deliver technical artifacts like MCP servers, sub-agents, and agent skills, provide white-glove deployment support in enterprise environments, and identify reusable deployment patterns to feed back to product and engineering teams."
That last part—"feed back to product and engineering teams"—is the true leverage of the FDE model. It means every on-site engagement is both a delivery for the client and a product discovery for the vendor. FDEs are the vendor's tentacles reaching into the market, bringing back samples of real-world needs.
IV. Is FDE the "PMF Paradigm of the Agent Era"? Three Reservations
By now, the judgment that "FDE is the PMF paradigm of the Agent era" sounds very convincing. But accepting this conclusion broadly ignores several real paradoxes.
Reservation 1: FDE might be solving the "PMF problem," or it might be "masking the PMF problem."
The original meaning of PMF is "product fits the market"—the product itself is the answer, and clients use it, renew it, and recommend it immediately.
The essence of the FDE model is "using human labor to bridge the gap between product and market." If a product requires a team of engineers on-site for six months to get running, strictly speaking, the product itself hasn't found PMF.
Alex Coqueiro, a senior analyst at Gartner, gave a stinging prediction in a recent report: by 2028, 70% of enterprises will be forced to abandon FDE-led Agent projects because "vendor costs are too high and internal capabilities for independent evolution are lacking."
He also pointed out a hidden failure mode: "If FDE workload does not decrease after multiple deployments, it is a signal that dependency rather than capability is being built. When a use case matures but investment doesn't drop, it means clients are paying consulting prices for operational capabilities they should own themselves."
This is the biggest risk of the FDE model: it could degenerate from a "product discovery mechanism" into "permanent labor filling." The reason the Palantir model succeeded was the "gravel road to paved highway" step—the specificity of client scenarios must eventually be distilled into the product. If this distillation step fails, FDE is just high-end outsourcing.
Reservation 2: Is this a "consulting firm disguised as a product company"?
Capital market judgment on this is also split.
Supporters believe the FDE model gives AI companies a "pre-deployment" moat: the earlier you send engineers into Fortune 500 companies, the earlier you control the entry point for enterprise AI workflows, and client migration costs will rise exponentially. OpenAI Deployment Company's official statement mentioned that partners "sponsoring over 2,000 enterprises globally" will become DeployCo's natural client pool—both a source of revenue and a feedback loop.
But critics point out that this model makes the financial profile of AI companies look more like a "consulting + software" hybrid. Palantir has long been undervalued in the secondary market partly because analysts use pure SaaS valuation frameworks (high margins, low labor) that don't fit. As OpenAI and Anthropic start hiring FDEs at scale, their margin structures, revenue per employee, and valuation multiples will be challenged.
Constellation Research analyst Larry Dignan's evaluation was more direct: OpenAI Deployment Company doesn't operate independently like IBM Consulting, which can integrate any model. "The chance of OpenAI Deployment Company using Anthropic is zero. OpenAI portrays its service department as a vertical integration advantage, but CIOs will see it through the lens of 'lock-in'."
In other words: what is a PMF paradigm for the vendor might be the eve of vendor lock-in for the client.
Reservation 3: FDE might be replaced by the tools they create.
This paradox is the most interesting. FDEs are expensive because they do a lot of "integration dirty work": field mapping, API interfacing, legacy system translation, prompt tuning, and building evaluation frameworks—precisely the types of work AI is best at automating.
Salesforce's practice with its Agentforce product shows that much of the "simple FAQ Agent deployment" work initially done by FDEs is being absorbed by the product itself; FDE work is migrating to higher abstraction layers—multi-agent architecture, MCP protocol design, voice Agents, and coding Agent orchestration.
In an April 2026 roundtable on FDEs hosted by South Park Commons in New York, several FDE heads reached a consensus: as models get stronger, the value of FDEs doesn't decrease, it increases—but the source of value changes. Low-level integration work is eaten by AI, and the core value of FDEs shifts to "judging which problems to solve at the client site and what to standardize."
This is a delicate balance. If AI tools evolve fast enough, the "integration leverage" of the FDE model will be compressed, leaving only product judgment and business consulting—then it truly becomes "high-end consulting." But if AI evolution hits a bottleneck, the complexity of integration will persist for many years, making FDE a long-term business.
V. The Meaning Varies for Different People
Back to the original question: Is FDE the PMF paradigm of the Agent era?
If I must make a judgment, I tend to state it this way: FDE is the "necessary intermediate state" for enterprise AI to move from demo to production in the Agent era, but it is not PMF itself—it is the method for finding PMF.
This statement has different meanings for different identities:
- For AI vendors: FDE is not a revenue business; it's a product discovery mechanism. If you treat it as a consulting business, you'll fall into a margin trap; only by continuously distilling on-site experience into reusable product capabilities—MCP servers, agent skills, evaluation frameworks, deployment templates—will FDE investment yield compound interest.
- For enterprise clients: The true value of FDE is not letting the vendor "build it for you," but "transferring capability to you in the process of building it." In the official statement of the Anthropic-FIS partnership, this sentence is key: "transfer knowledge so FIS can build and scale additional agents independently over time." If there's no such exit mechanism in the contract, the FDE model is a gentle lock-in.
- For engineers: This is the rarest skill set of 2026—technical depth, understanding of client context, and business judgment. Google's listed FDE salary range is $127k to $265k base, with senior packages averaging $238k and top-tier approaching $400k. Moreover, this budget comes from client expansion spending, not internal R&D headcount, making it counter-cyclical during layoff periods.
- For investors: Using a pure SaaS valuation framework for FDE-driven AI companies will be misleading. What needs to be watched is not current margins, but the speed of "turning gravel roads into paved highways"—how much the product's reusable capability improves after each on-site engagement. It took Palantir nearly twenty years for the market to understand this; OpenAI and Anthropic won't have that much patience.
Conclusion: Paradigms Don't Announce Their Own Birth
The term PMF was first proposed by Marc Andreessen in 2007, and his criterion was very simple: "You don't need to explain it, you just know you've found it."—users start flooding in, the product is in short supply, and the system is constantly overloaded.
By this standard, the AI enterprise market in May 2026 has the "embryo of PMF" but not yet the "victory of PMF." The three companies betting on FDE simultaneously is less about declaring a paradigm victory and more about admitting a fact: before Agents become true "software above software," we need people—on-site people who understand both the client and the model—to walk those unpaved roads one by one.
Perhaps the true PMF paradigm will wait until the roads walked by FDEs are numerous and clear enough that Agents can run on them themselves—at that point, this discussion about FDEs will become a footnote of an era.
But in 2026, everyone is still on the road.
Data and cases in this article are from official announcements by OpenAI, Anthropic, Google, and FIS, as well as public reports from The Information, Pragmatic Engineer, Constellation Research, CIO Magazine, and Gartner, with data current as of May 2026.
