Back to writing

What Forward-Deployed Engineering Teaches You

What three and a half years as a forward-deployed engineer at Gecko Robotics taught me about building AI products.

When I told people my title was “Forward Deployed Software Engineer,” the usual response was a polite nod and a subject change. Fair enough. It sounds like a military posting. In a way, it is: you’re an engineer stationed at the customer, building product from inside their problem.

Now that OpenAI, Anthropic, and many other AI companies hire forward-deployed engineers to get agents into production, the model Palantir pioneered has become a standard way to ship AI. I did the job for three and a half years at Gecko Robotics, across power plants, refineries, and paper mills. These are the lessons that stuck.

You can’t spec what you haven’t seen

Asset Manager, the product behind Gecko’s first SaaS deal, didn’t come from a product-requirements document. It came from watching how reliability engineers actually made capital decisions: spreadsheets, tribal knowledge, and PDFs of inspection reports. Seeing that up close made it obvious the data from our robots could replace guesswork with evidence. You could only see it from next to the user, not from headquarters.

That is the core trade of the job: you give up the clean abstractions of platform work and get an accurate picture of the problem in exchange.

Speed is a feature of proximity

When the person who feels the pain sits across the table, iteration cycles collapse from quarters to days. Ship something rough on Tuesday, watch it fail in an interesting way on Wednesday, ship the fix on Thursday. Often the feedback is a facial expression, not a dashboard metric.

This is why the FDE model fits AI so well. Agents are too new, and too dependent on messy organizational context, to be built from spec sheets. The teams shipping working agents are the ones embedded close enough to watch them fail in the wild.

The prototype is a promise

The FDE failure mode is the permanent hack: a demo that works for one customer and calcifies into unmaintainable glue. The discipline that separates forward-deployed engineering from consulting is knowing which prototypes are throwaway and which are the first draft of a product. Then comes the harder second job of generalizing. Cantilever AI started as work for specific power customers; it became a platform because we kept asking which parts of the solution were about this plant and which were about every plant.

Empathy compounds like technical skill

The most useful thing I took from those years is a reflex. Before writing code, I ask: who touches this, what were they doing right before, what happens to them if it’s wrong? Now that agents act on people’s behalf, those questions are part of the engineering.

Different industries ask different questions, but the method held up everywhere I used it: sit close to the problem, ship, watch, generalize.