Really useful glossary, and I especially appreciate the attempt to separate the model, harness, scaffold, tools, context, and training loop.
One concern I have is with standardizing on “scaffolding” as the term for the behavior-defining layer around the model. To me, “scaffolding” already feels quite overloaded, specifically in software projects. It is used extensively in code generation, and more narrowly by framework generators to create routes, controllers, models, migrations, components, tests, and other boilerplate structures.
I wonder whether “rigging” might be a better candidate. It still relates naturally to the metaphor of a harness, but is less already-standardized in adjacent software and AI contexts. In the horse/cart sense, rigging is the arrangement that attaches the harness to the thing being pulled and makes the whole setup usable. That feels close to what this layer does for an agent: system prompt, tool descriptions, response formats, memory/context rules, and other configurations that connect the model to the operational environment in a particular way.