MTTR-A: Measuring Cognitive Recovery Latency in Multi-Agent Systems (arxiv.org)

A new metric called MTTR-A (Mean Time-to-Recovery for Agentic Systems) has been introduced to address the challenges of cognitive recovery latency in multi-agent systems (MAS) that utilize large language models. Unlike traditional reliability assessments that focus on infrastructure faults, MTTR-A quantifies the time it takes for these systems to detect reasoning drift and regain coherent operation. This is crucial as cognitive failures are becoming more common in MAS, potentially undermining their effectiveness. Accompanying metrics, such as Mean Time Between Failures (MTBF) and Normalized Recovery Ratio (NRR), further enhance the understanding of cognitive dependability. The significance of MTTR-A lies in its ability to provide a quantitative foundation for assessing runtime cognitive reliability in distributed systems. By employing LangGraph-based benchmarks that simulate reasoning drift and reflex recovery, the research showcases measurable recovery behaviors across various strategies. This innovative approach not only advances the field of multi-agent system reliability but also lays the groundwork for improving the performance of AI systems in real-world applications, ensuring they can recover effectively from cognitive disruptions.