When to Stop Using Forward Deployed Engineers (FDE)

Key takeaways

1. Date / source. August 11, 2026, long-form article on X by Jesse Zhang, CEO of Decagon.
2. Key framing. the rule fits in one sentence — « The trap is not starting. It's not stopping. » ### The four-question test Value that transfers beyond the article, independent of Decagon — quarterly-review format: | Question | What it discriminates | |---|---| | Is the bespoke work in the customer's environment or in the holes in your product? | legitimacy of custom work | | Is the last mile irreducible or simply unbuilt? | fate vs. debt | | Are your FDEs discovering or absorbing? | discovery vs. amortization | | What was put into the product the last time an FDE came back from the field? | proof, not intent | The fourth is the only verifiable one: the first three answer themselves favorably, this one demands an artifact. ### The mechanics of drift Keeping the FDEs is easier sprint by sprint taken in isolation: the FDE lets a company « avoid every hard product trade-off » — one never decides what the product does, which of two customer requests wins, where the configuration surface stops. « No one has to say no to anyone. » The drift comes from no single bad decision but from the repeated absence of one. Associated cost: « Every bespoke fix in the field is a product decision you chose not to make. » ### FDE ≠ implementation | | FDE | Implementation | |---|---|---| | Object | discovery of an unknown spec | execution against a known spec | | Example | sitting in the room, watching the product break | « building the integration into their ticketing system » | | Expected output | product primitives | a customer deliverable | « Lumping the two under one title is how a company convinces itself that a growing services org is a product investment. » Inventory test: count, within a team labeled FDE, the share of work that is actually implementation. Zhang adds that this is the half models are absorbing — « a good part of what an implementation team did in 2023 becomes something the product does itself ». ### The Palantir genealogy, a textbook case of "services → product" Bespoke Gotham deployments (CIA, NSA, military intelligence, mid-2000s) → encoding the problems encountered into platform primitives (ontology, object models, permissions, workflow engines, provenance tracing) → Foundry, commercially sellable → Apollo, AIP → standardization, gross margin in the 80% range, shift to account-based selling, FDEs reabsorbed into core engineering. Two details that make the case: Palantir carried the « glorified consultancy » criticism for nearly twenty years — Lonsdale acknowledging the observation was accurate — and turned down contracts where the customer just wanted « Accenture with better software ». Line worth keeping: « The FDE team wasn't the business model. It was how you built the right product. » ### Counter-test: the same indicator, two readings In May 2026, [[mollick-roon-asi-consulting-forward-deployed-engineering-2026-05-10]] argued we'd know the labs believed in ASI the day they dissolved their FDE teams — and observed that they were hiring them instead. Zhang describes, three months later, the reverse movement at startup scale: the product eats the deployment work. Both texts use the same indicator — the size of the FDE org as a measure of what the product still can't do — one to cast doubt on a narrative, the other to claim progress. They don't contradict each other: Zhang confirms the FDE-hiring trend and himself notes that Anthropic and OpenAI have built deployment arms modeled on Palantir. ### Why now, and the expiration date In 2015, building a SaaS CRM required no workflow discovery — twenty years of practice had already defined what a pipeline, a stage, a lead handoff was. In 2026, an AI agent for accounting has no established workflow, « because literally no one has ever used one ». Corollary: the customer can't say what they want, because the thing they'd want doesn't have a shape yet. The FDE is therefore justified only by the category's novelty; as soon as the shape stabilizes, the justification falls away. ### The two figures, and how to cite them | Stated figure | What's missing | Acceptable usage | |---|---|---| | « Two-thirds of deployment work is now done autonomously via Duet » | definition of the denominator (hours? tickets? steps?), scope, period, method | « Decagon states », never « Decagon measured » | | « A few days on average to launch the first AOP, even for large banks, airlines, telcos » | the AOP acronym is not spelled out; no starting point for the clock, no sample size | a dated commercial claim | These are executive assertions on X, made the day he's defending his model: the source and the date are part of the figure. ### The two constants heard from enterprise customers 1. Iteration speed — « Shipping an AI agent isn't a one-shot; it has to be tuned and updated continuously. If every adjustment requires engineering, it'll be far too slow and expensive to scale. » Structural argument against the permanent FDE: a question of loop latency, not margin. 2. Vendor lock-in — « Given organizations' experience with SaaS, no one wants to be locked into a vendor and dependent on its resources. » An FDE org is a dependency, from the customer's point of view. ### The declared trade-off and its limit Decagon says it chose not to hack fixes in the field when that would have been faster, and instead to turn escalations into requirements rather than patches — « which takes time in the short term ». Formula worth keeping in architecture review: escalation → requirement, not escalation → patch. Honest counterpart: « very few startups can sign the eight-figure deals Palantir landed from the start », which makes the FDE economics even less sustainable for them. Limit of applicability stated by the author: the product-led approach holds for Decagon because customer service is « high-volume, repeatable, and decomposable ». Those three adjectives are the condition — a low-volume, non-repeatable, non-decomposable domain doesn't tip the same way. Don't carry the conclusion without carrying the condition.