FLASH: A Workflow Automation Agent for Diagnosing Recurring Incidents

Recurring incidents, typically raised by system monitors, often occur repeatedly, demanding significant human effort for troubleshooting. Automating the diagnosis process for these recurring incidents is crucial for minimizing service downtime, reducing customer impact, and decreasing manual labor. While recent agent approaches based on Large Language Models (LLMs) have demonstrated effectiveness in handling complex tasks requiring multiple logical steps, they still suffer from the reliability issue due to a lack of specific diagnostic knowledge. To enhance diagnostic reliability, we propose a workFLow Automation agent with Status supervision and Hindsight integration (FLASH), which significantly improves diagnostic accuracy by incorporating status supervision to break down the complex instructions into manageable pieces aligned with identified status. Moreover, we generate hindsight using LLMs from past failure experiences, progressively enhancing diagnostic reliability for subsequent incidents. We conduct extensive study over 250 production incidents from Microsoft in five different workflow automation scenarios. The results reveal that our FLASH agent approach outperforms state-of-the-art agent models by an average of 13.2% in terms of accuracy. These compelling results underscore the viability of automating the diagnostic process for recurring incidents.