The model won't devour everything: three primitives of agent systems

The analysis derives from five evidence streams and presents eight testable predictions with operational acceptance criteria to be evaluated through late 2028.

The framework's central argument is that AI models will not 'devour everything' but will continue replacing acceptance-executable, non-physical components. The model proposes that runtime will thin out, with its judgment half absorbed by base-model iteration while enforcement half persists as necessary boundaries.