AAM-V1: Topological Telemetry and Evidence-Based Runtime for Long-Horizon Autonomous Agents

Abstract[Human-Readable Summary] Autonomous agents based on Large Language Models (LLMs) are highly vulnerable to recursive collapse (“rabbit holes”)—an infinite loopof ineffective actions combined with a false inflation of internal confidence. This paperpresents the Artsybashev Analysis Method (AAM-V1), a non-invasive middleware designed to ensure agent reliability. By shifting the focus from semantic analysis to thespectral-topological analysis of latent space geometry and empirical data verification,AAM-V1 effectively interrupts cognitive looping. Experimental validation demonstratesa 65–77% reduction in looped sessions, massive data center compute savings, and theprevention of Recurrent Data Poisoning in RLAIF pipelines.Keywords: Autonomous Agents, LLM, Topological Telemetry, Runtime Reliability.Canonical ID: AAM-V1_ARTSYBASHEV_UA_KHARKIV_AIANALYSIS