Long-horizon autonomous agents frequently enter recurrent, low-productivity cycles prior to triggering explicit system failures, such as execution timeouts or token exhaustion. This paper introduces AAM-V1, an architecture-agnostic runtime telemetry middleware designed to predictively detect these silent degradations. We hypothesize that trajectory dimensional collapse in the agent's latent state-space acts as a robust, early-warning signal for recurrent failure. The proposed framework evaluates sequences using a minimal set of computable runtime metrics: Participation Ratio (PR), Recurrence Determinism (DET), and Spatial Mobility. Through passive replay-backtesting on historical logs from SWE-agents, robotics planners, and swarm simulations, we demonstrate that AAM-V1 yields a measurable Predictive Horizon Gain (PHG) while resisting false positives through a formalized Separation Lemma. This work shifts the observability paradigm from step-level syntactic correctness to continuous geometric trajectory evaluation, providing a critical safety and recoverability layer for autonomous deployment.
ANDRII ARTSYBASHEV (2026) studied this question.