Sit with the desk, or hand over an outcome.

A forward-deployed engineer is there because the use case is tangled up in how a team already works. The packet has exceptions everyone handles from memory. The approval is real and also informal. The system of record is the one people actually update, which may not be the one the diagram names. Those facts are learned by sitting with the desk, not by reading a requirements document at a distance.

The engineers are E31-certified: trained and qualified by Element 31 on Czar, Substrate, and Agency before the engagement. On site they compose the agent with your operators, mark the step a person must own, name the calls, and publish a version against the work the desk already has. They stay through the point where your people change that version without them.

It fits when the outcome depends on the room. A claims desk, a review cell, a shift that already has a way of working — the engagement succeeds when that way of working is what the agent does, and when the people on the shift can alter it.

An in-house project starts from an outcome you can describe. What should be true when the engineers are done: a corpus the staff can question, a review that drafts and waits, an application your customers run under your name. The same training, a different location. The work happens with Element 31, against that scope, on the appliance you have.

What comes back is the running use case. The published version. The collections it may read. The systems it may call. The role that approves. Notes written for the operators who will change it next. Your team can run it, and they can publish the following version on Agency.

It fits when you can name done, and when you want the first version built by people who already know the stack, before you stand up that skill entirely inside the organization.