The author argues that describing modern, self-hosted long-lived agent systems merely as "customized assistants" or "memory plus persona" is no longer technically sufficient. These older frames reduce complex runtime behaviors to simple style and retrieval, ignoring critical distinctions in continuity, provenance, and authority discipline.
- The proposed framework separates candidate generation from admissible action within governed execution surfaces.
- A key principle is that model confidence does not equate to authority or permission for action.
- Internal bounded tests using a typed-recovery benchmark with 40 distinct cases showed zero wrong-authority accepts in the governed lane.
- The author identifies specific engineering questions regarding state admissibility, contamination prevention, and recovery discipline.
This shift in terminology is necessary because continuity and governance materially alter how these systems judge and execute actions, requiring more rigorous technical evaluation than raw fluency.