Why agents need forward deployed engineers to scale
Bob McGrew traces how Palantir’s forward deployed engineer model became the default playbook for AI agent startups facing massive product discovery and no incumbent categories. The strategy pairs embedded “echo” analysts with rapid-prototyping “delta” engineers to deliver outcome-based value, grow contract size, and feed generalized platform roadmaps rather than bespoke consulting. The result: do things that don’t scale—at scale—while turning gravel-road prototypes into superhighways for the next 5–10 customers.
Points clés
- Bob McGrew’s background includes early roles at PayPal and Palantir and serving as chief research officer at OpenAI, where he helped lead ChatGPT, GPT-4, and the o1 reasoning model; he has also joined the US Army Reserve (Detachment 2011) as a lieutenant colonel.
- A forward deployed engineer (FDE) sits with the customer to bridge the gap between what the product does and what the customer needs, delivering a concrete outcome rather than a generic install.
- Palantir’s model paired “echo” teams (embedded analysts/account owners) with “delta” teams (deployed engineers who prototype fast and “eat pain”) to land on a high-impact use case, then expand.
- On the YC job board, over 100 YC startups are now hiring for forward deployed engineers—up from essentially none three years ago—underscoring the model’s adoption among AI agent companies.
- The FDE strategy flips classic product-market-fit scaling: instead of minimizing per-customer work, you increase the value of outcomes and drive contract size up while productizing common patterns.
- Pricing is outcome-based, not seats or usage; early deployments may be unprofitable, but margins improve as product leverage rises and teams “earn the right” to tackle higher-value problems.
- Product teams must generalize field prototypes into reusable platforms (e.g., Palantir’s ontology of objects, properties, and links), avoiding over-specialized features that lock to a single customer.
- Organizational discipline is crucial: secure executive backing to target a CEO’s top priorities, overcome IT friction, and avoid devolving into consulting that builds what users ask for instead of what moves the business.
- Demo-driven development (e.g., Palantir’s “stop a terrorist plot” end-to-end demo) forces features to cohere into workflows that customers viscerally want, accelerating adoption and internal alignment.
- AI agent markets favor FDEs because there is no incumbent product and heterogeneity is high; capabilities are racing ahead (e.g., from GPT-4o in April 2024 to o3 in April 2025), but adoption lags—creating opportunity for embedded teams to translate breakthroughs into durable enterprise outcomes.
À retenir
Want to make AI agents useful instead of just impressive? Sit with the customer, pick one of the CEO’s top five problems, and ship a prototype that moves a KPI anyone can count—then keep turning that gravel road into a highway. Hire rebels (echo) and rapid-builders (delta), price the outcome (not the seats), and track two things: growing contract value and rising product leverage. And no, a steak dinner with Don Draper won’t replace executive buy-in—but it might replace your runway, so maybe keep the tab modest.
Sources
Quiz sur la vidéo: 5 questions
