Abstract
Reliability failures in LLM-based multi-agent systems are often cognitive rather than infrastructural: agents may keep running while reasoning drifts, memories desynchronize, or coordination degrades. This paper introduces MTTR-A, Mean Time-to-Recovery for Agentic Systems, as a runtime reliability metric for measuring cognitive recovery latency in multi-agent orchestration. The framework adapts classical dependability metrics such as MTTR and MTBF to distributed reasoning, defines a normalized recovery ratio (NRR), and links recovery latency to long-run cognitive uptime. A LangGraph-based benchmark with simulated reasoning drift and reflex recovery evaluates recovery behavior across multiple reflex strategies, establishing a practical foundation for quantitative runtime dependability in agentic AI systems.