Autonomous Agents as First-Class Operating System Entities: Formal Architecture, Deterministic Protocols, and Self-Regulating Runtimes

Modern multi-agent architectures predominantly treat autonomous agents driven by Large Language Models (LLMs) as unmanaged application-level loops interacting through stateless, unbounded prompts. In long-running computing environments, this abstraction mismatch induces severe system-level pathologies, including nondeterministic context drift, unconstrained heap expansion, event-driven thread starvation, and orphaned subprocess leakage. In this paper, we formalize, implement, and empirically validate the paradigm of Autonomous Agents as First-Class Operating System (OS) Entities. Under this paradigm, an autonomous agent is elevated from an ephemeral prompt string to a sovereign execution unit governed directly by operating system process abstractions: possessing a dedicated Process Identifier (PID), an isolated Pseudo-Terminal (PTY) communication channel, hardware telemetry self-awareness, and explicit lifecycle semantics. We make four primary theoretical and empirical contributions: (1) a formal model defining the Four Axioms of a First-Class Agentic Entity, establishing invariants for process isolation and deterministic execution, accompanied by formal mathematical proofs of Bounded Heap Invariants (via Foster-Lyapunov drift criteria) and Deadlock-Freedom (via Petri Net reachability); (2) the Autonomous Signal & Context Bridge (ASCB) protocol, a 5-phase deterministic state machine that eliminates hallucinatory divergence through verifiable cryptographic handshakes; (3) an adaptive, kernel-level telemetry mesh featuring LeakSentinel, an Ordinary Least Squares (OLS) linear regression engine evaluating memory growth rates with the Pearson coefficient of determination (R² ≥ 0.70), coupled with a three-phase automaton for autonomous idle memory reclamation; and (4) an extensive suite of empirical benchmarks on a physical multi-core testbed, including a continuous 24-hour soak test, demonstrating a 7.28× reduction in concurrent request latency (from 4,019 ms to 552 ms for 60-burst workloads) via Singleflight request coalescing, an 85% curtailment of background OS probing overhead, and absolute zero subprocess orphanage under abrupt process termination.