Managed Autonomy at Runtime: Gear-Based Safety and Governance for Single- and Multi-Agent Cyber-Physical Systems
06:00 · July 2, 2026 · arXiv cs.AI RSS
Autonomous agents, whether LLM-driven software agents or robotic physical agents, face a common class of failure modes when operating without continuous human oversight: safety violations from unverified actions, behavioral instability from unconstrained loops, and continuity loss from unhandled error states. We develop \system{}, a discrete-time control system that combines five execution gears (\Gobs{}, \Gsug{}, \Gplan{}, \Gexec{}, \Gint{}) with utility-gated dispatch and event-driven fallback. For the single-agent case, we prove monotonic stability, execution safety, eventual stabilization, fallback completeness, and equivalence to a gear-constrained Markov decision process. For multi-agent cyber-physical systems (CPS), we apply the established \smart{} managed-autonomy lifecycle and map runtime evidence into its four governance states (\Stable{}/\Meta{}/\Assisted{}/\Regulated{}). Consensus gating, swarm-level Lyapunov analysis, per-agent gear authority, and rendezvous control provide distributed safety and stability guarantees, including zero collision under the stated assumptions. We evaluate the resulting runtime on a three-agent UR5 robotic assembly cell using fault magnitudes calibrated from the NIST \emph{Degradation Measurement of Robot Arm Position Accuracy} dataset across 10,000 Monte Carlo episodes. It achieves a 99.6\% anomaly detection rate versus 2.1\% for the single-agent baseline, reduces detection latency by $3.5\times$, and supplies a formal physical-workspace safety certificate. The execution gears act as micro-level permissions beneath the \smart{} runtime governance states, separating action control from autonomy governance.
Summary
Autonomous agents in both software and robotic settings encounter recurring failure modes when left without continuous oversight, including unverified actions that produce safety violations, oscillatory or looping behavior, and abrupt loss of operational continuity. EntropyRuntime addresses these issues through a discrete-time control architecture that interleaves five execution gears—Observe, Suggest, Plan, Execute, and Integrate—with utility-gated dispatch and event-driven fallback. Each gear restricts the permissible action scope at runtime, while the utility gate blocks any candidate whose expected value falls below a defined threshold.
For single-agent operation the framework supplies formal proofs of monotonic stability, execution safety (no negative-utility action is ever dispatched), eventual stabilization, and fallback completeness. It also establishes equivalence to a gear-constrained Markov decision process, allowing standard dynamic-programming techniques to be applied under the added constraints. These guarantees are obtained by treating gear transitions as a finite-state process whose invariants are preserved across cycles.
In multi-agent cyber-physical systems the same gear mechanism is placed beneath the SMARt managed-autonomy lifecycle. Runtime evidence is mapped into four governance states—Stable, Meta-Cognitive, Assisted, and Regulated—while a consensus utility gate, swarm-level Lyapunov analysis, per-agent gear authority, and rendezvous control together enforce distributed safety properties, including zero workspace collisions under the stated assumptions. The gears therefore function as micro-level permissions that realize the higher-level governance decisions without redefining the lifecycle itself.
The architecture was evaluated on a three-agent UR5 robotic assembly cell whose fault profiles were calibrated from the NIST robot-arm degradation dataset. Across 10,000 Monte Carlo episodes the system recorded a 99.6 % anomaly detection rate (against 2.1 % for the single-agent baseline), reduced detection latency by a factor of 3.5, and supplied a formal physical-workspace safety certificate. Ablation experiments confirmed that each component—gear abstraction, utility gating, and event-driven fallback—is necessary for the observed safety and stability properties.
Why it matters
This research is highly relevant to Dutch AI researchers and the high-tech systems sector (e.g., Brainport Eindhoven) as it provides formal safety and governance frameworks for autonomous cyber-physical systems. Its focus on verifiable safety and runtime governance aligns strongly with EU AI Act requirements for high-risk AI applications.
