Palantir’s original FDE structure had two roles. The Echo was an embedded analyst with deep domain expertise (a former military officer in a defense engagement, a former clinician in healthcare) responsible for understanding the customer’s domain in its bones and managing the relationship. The Delta was the deployed engineer, responsible for taking the problem the Echo identified and building a working system on a short timeline.
Most AI startups today don’t have an Echo team. They have FDEs who do both jobs at once. In the short term the consolidation works. But the key insight in the split isn’t the org structure. It’s that the role contains two fundamentally different kinds of work, and the FDE has to be deliberate about which one they’re doing at any given moment. Discovery means listening, observing, and building a model of the customer’s reality. Build means writing code, integrating systems, and shipping. Blur them, by building before you’ve understood or still listening when you should be shipping, and the engagement loses momentum.
The Echo mode: find the heretic
Shyam Sankar’s description of the Echo profile is worth holding onto: not a domain expert in the standard sense but a domain rebel. A heretic. Someone who came from the domain, understands it deeply, and has concluded that the way it’s currently done is inadequate. A satisfied practitioner will tell you the current way is fine and your system is unnecessary. A heretic will tell you exactly what’s broken, why every previous fix failed, and what would have to change for a new attempt to succeed.
Three Echo skills matter most:
Ask one more question after the polite stopping point. Most discovery interviews end too early. The question after the polite question (”What’s the part of your work you wish someone outside this team understood?”) usually surfaces the real problem instead of the stated one.
Suppress the pitching reflex. Most vendor-side employees, hearing a customer describe a problem, immediately map it onto what their product can do. The mapping is reflexive and usually wrong. Rob Fitzpatrick’s The Mom Test is the operating manual here: talk about their life, not your idea; ask about specifics in the past, not opinions about the future; talk less and listen more.
Map the politics. Learn early who can keep the project alive when it hits friction, who can kill it, who has operational responsibility, and who holds the latent veto. The map is rarely the org chart.
The Delta mode: eat pain without losing energy
The Delta is a different animal from the canonical software craftsman, the engineer whose deepest pleasure is refactoring a codebase into an architecture that will still be maintainable in fifteen years. The first version of anything the Delta builds will probably be thrown away. The skill that matters is writing code that ships, today, against a moving target, with imperfect information. Bob McGrew calls this eating pain.
The second Delta skill is knowing when to push back on the customer’s preferred solution. The best interview question for it: describe a project where the customer asked for X and you built Y. A good answer is a coherent story about why X was the stated request, what about it didn’t solve the problem, how you discovered that, and how you persuaded the customer that Y was better. A bad answer makes the customer sound foolish or the candidate sound rigid. The work is negotiated judgment.
The split in a single engagement
One engagement makes this concrete. A national wholesale distributor of industrial parts stated its problem as “we want AI to help our sales reps quote orders faster.” The first ten days were almost pure Echo: the FDE sat next to sales reps for full shifts and watched the work. The reps weren’t slow at quoting. They were slow at finding ground truth across three back-office platforms whose inventory data disagreed. The domain rebel turned out to be a sales operations manager who had spent three years quietly maintaining a personal spreadsheet that reconciled the platforms better than any of them did on their own.
Ten days of Echo work produced no code, but it did produce the diagnosis that drove every Delta decision for the next ten weeks. The first slice, a reconciliation pass that ran her heuristics in code, shipped on day twenty-two. The reps loved it, and the eventual quoting agent was built on top of it. Pitching a quoting agent in week one, or stretching the listening phase into week four, would have produced a worse engagement either way.
The rhythm generalizes. The first week is mostly Echo, with as little Delta as possible. The middle weeks weight heavily toward Delta. The last weeks are mostly Delta, with Echo used to validate that what’s being built will land. Listen carefully in one mode, ship aggressively in the other, and be honest with yourself about which one you’re in.
Want the full playbook, with the templates, the interview transcripts, and the week-one checklist? It’s all in my book Forward Deployed AI Engineering: A Working Guide to the Hottest Job in Software, available on Amazon: [Amazon link]