Measuring Multi-Turn Consistency and Failure Modes in LLM Agents
September 25, 2026
An evaluation of 84,540 trajectories across 8 model families quantifies how agents drift toward inconsistency during 20-step multi-turn interactions. The study uses survival analysis and a seven-category failure-rationale taxonomy to identify how social visibility and persona stressors impact agent reliability.
HOW THIS AFFECTS YOU
●
builderYou should account for temporal drift and decision inconsistency when deploying agents in long-running, multi-step workflows.
●
researcherThe failure-rationale taxonomy provides a structured way to categorize agentic errors.