Discrete State, Continuous Agency: An Engineering Profile for Intermittent AI Executives

Persistent artificial systems increasingly combine fast controllers, durable services, asynchronous observations, external tools, and generative reasoning components. Their continuity cannot be inferred from the uptime of any one model process. This white paper defines an engineering profile for systems in which an executive reasoning model may be idle, interrupted, restarted, or replaced while the world and lower-level processes continue to evolve. The profile couples five elements: bounded local control, an authoritative operational record with explicit provenance and partial order, epoch-fenced commands, durable handling of external effects, and a task-conditioned reconstruction that must precede renewed proposal access. It separates runtime, control, task, interpretive, narrative, and subjective continuity, and addresses only the first four. A formal summary-sufficiency condition shows why narrative or averaged state cannot serve as authoritative memory across unrestricted future goals. Concrete failure boundaries are specified for delayed and reordered messages, stale reasoning, lost acknowledgments, duplicated effects, conflicting controllers, expired authority, interacting control loops, and goal changes during interruption. The component technologies are established in robotics, distributed systems, control, workflow orchestration, and agent runtimes. The contribution is a vendor-neutral implementation synthesis for contract-constrained generative executives supervising live, asynchronously changing, potentially non-rewindable systems. A small two-carrier interrupted-delivery benchmark is proposed to distinguish this profile from strong conventional hierarchical and flat event-state alternatives. No empirical results are reported.