Multi-agent systems (MAS) often suffer from instability driven by non-stationarity, coordination failure, and uncontrolled feedback loops, frequently leading to performance degradation or total system collapse. In this concept paper, we introduce Dynamic Agent Stability Control (DASC), a framework that models and mitigates instability through a dynamically activated supervisory agent. We define a quantitative, collapse-aware stability metric based on reward dynamics and propose a threshold-based mechanism to trigger intervention strictly when instability is detected. Unlike traditional centralized control or always-on hierarchical approaches, DASC operates adaptively at runtime. This minimizes unnecessary interference during stable execution while significantly improving system robustness during chaotic episodes. This preprint outlines the theoretical foundation, mathematical formulation, and proposed experimental setup for evaluating DASC against unsupervised and statically supervised multi-agent baselines.
Ujjwal Rai (Sun,) studied this question.