AAM-V1: Spectral-Topological Runtime Telemetry for Early Detection of Trajectory Stagnation in Long-Horizon Autonomous Agents

Long-horizon autonomous agents, particularly those based on Large Language Models (LLMs) and iterative planners, are prone to latent performance degradation where the system enters low-diversity, recurrent state cycles. These "attractor locking" events often consume significant computational resources before triggering standard timeout-based interventions. This paper presents AAM-V1, a lightweight runtime telemetry framework that detects trajectory stagnation by monitoring three spectral-topological metrics: Participation Ratio (PR), Recurrence Determinism (DET), and Centroid Mobility (M). We introduce the Separation Lemma to distinguish productive task specialization from pathological stagnation. Unlike existing heuristics, AAM-V1 provides a positive Predictive Horizon Gain (PHG) with minimal computational overhead. We provide an asymptotic complexity analysis and a reference implementation to ensure reproducibility in production environments. Independent Researcher, Kharkiv, Ukraine или AAM-V1ARTSYBASHEVUAKHARKIVAIANALYSIS