New System Enhances Safety and Governance for Autonomous Agents

A new discrete-time control system has been developed to address common failure modes in autonomous agents, particularly when operating without continuous human oversight. These failure modes include safety violations from unverified actions, behavioral instability from unconstrained loops, and continuity loss from unhandled errors. The proposed system integrates five distinct execution "gears" – observation, suggestion, planning, execution, and intervention – with utility-gated dispatch and event-driven fallback mechanisms. For single-agent systems, the research proves properties like monotonic stability, execution safety, and eventual stabilization. When extended to multi-agent cyber-physical systems (CPS), the system applies a managed-autonomy lifecycle, mapping runtime evidence into four governance states. Distributed safety and stability are ensured through consensus gating, swarm-level Lyapunov analysis, per-agent gear authority, and rendezvous control, including a formal physical-workspace safety certificate for zero collisions under specified assumptions. Evaluation on a robotic assembly cell demonstrated a 99.6% anomaly detection rate, a 3.5x reduction in detection latency, and significantly improved safety compared to baselines.