Boundary Work, Not Castles

A language model transforms text into text. Nothing of value has happened until that text is wired into systems that act, and the wiring has a name older than the models: it is articulation work in Strauss's sense — the labour of fitting parts of a division of labour together — and it has the property articulation work always has. It is invisible, it is low-status, and it is therefore under-resourced at exactly the point where the value is created.

This paper contrasts two architectures for doing that work. One internalises composition inside a product: the agent, the framework — the castle. The other externalises composition to mechanisms that already exist: the pipe, the file, the exit code — the plumbing. The difference is not taste. In Perrow's terms the castle is interactively complex and tightly coupled where the pipeline is linear and loosely coupled; in observability terms the castle's intermediates are rendered where the pipeline's are returned, which fixes which levels of delegation an operator can occupy. We give a mechanism for why castles are built anyway — trained abstraction, the second-system effect, a demo gradient, and the economics of invisible work — and we state the conditions under which the castle is nevertheless the right structure, because a dichotomy without a failure mode is a slogan. A staffing consequence follows: the people fitted to boundary work already exist, hold operational job titles, and are not, in the main, the people currently assigned to it.

A case study shows the externalised architecture is constructible in one small tool; competing interests are declared. The claims are stated so that they can be tested, and Section 9 says how they would lose.