Deployment Engineering, In Practice

What the job actually is,
from inside the engagement.

Field notes on deployment-engineering practice — discovery, scoping, and customer judgment calls.

Field Logloading…

Refusal as Engineering Gate — The three diagnostic moments where an FDE's highest-leverage act is saying 'don't build this yet'—before premature engineering locks in failure.

Title as Encoding — When you see 'Forward Deployed Engineer' versus 'Solutions Engineer,' you're not reading a job description. You're reading three encoded decisions: comp bracket, risk allocation, and career structure.

Context Is the Integration — MCP turned FDE glue code into a durable asset the customer owns. The work that's left — and that now matters most — is curating what each connector puts in front of the model.

The Harness Moves the Number — Same 4B model, same task, a 60-to-82% accuracy swing from harness design alone. The eval harness is the FDE's real deliverable — and the acceptance contract with the customer.

The Handoff Is the Product — Air Canada, Klarna and Cursor learned the same lesson in public: the path from the bot to a human isn't the failure mode of an AI deployment. It's the feature that makes it safe to ship.

Prompt Injection Is a Permissions Bug — Assume the model will be tricked. The FDE's job is to decide, during scoping, what a tricked agent can reach — and to make sure it can't reach all three legs of the lethal trifecta at once.

Dry Run Before Autonomy — "Don't touch production" in a prompt is a request. A read-only credential is a control. Agent tool loops need dry runs, idempotency and scoped access before they need autonomy.

Rollback Is the Feature — The most instructive AI incidents of 2025 weren't fixed by better models. They were fixed by how fast someone could get back to the last known good state.

The Title Is The Tell — Fragmented FDE titles are not confusion — they are price discovery for deployment risk.

Not Yet Is the Deliverable — The best FDEs aren't measured by what they ship — they're measured by what they refuse to build until the data proves it's ready.

The Roadmap Is Not a Wish List — Internal advocacy is not lobbying. It is turning field pain into an eval product wants to ship.

Code Is Just a Receipt — In forward-deployed engineering, discovery isn't a prelude to the engagement. It IS the engagement. The code that follows is just the receipt.

The Code You Don't Write — The real product an FDE ships is a "don't build" list. Everything else is evidence for why each item belongs on it.

Lost Before You Type — Every failed engagement was lost in the first two weeks. The code just hadn't been written yet.

You're Not Deploying, You're Correcting — Your highest-output work isn't what you ship to the customer. It's what you ship back to your own product team.

Your Real Job Is Roadmap — The FDE's highest-leverage output isn't code deployed — it's field signal that changes the roadmap.

What the Job Selects For — The FDE role doesn't select for people skills. It selects for the ability to know the customer is wrong and say nothing — for three weeks.