LLM Based Agentic Framework to Assist with IT Incidents

When an incident or alert arises for a service, it often requires a significant amount of time for the on-call personnel to identify the root cause of the issue and subsequently work toward its resolution. In certain instances, the issue may also be linked to other dependent services, making it more difficult for an on-call individual to pinpoint the problem. This identification process can be extremely time-consuming and presents considerable challenges for the on-call engineer. In this paper, we will discuss a solution designed to assist on-call personnel in reducing Mean Time to Recovery (MTTR) through an agentic framework, which functions as a conversational AIOps Incident Management bot. In this paper, we will discuss our solutions to challenges such as recommending resources from various sources, utilizing features related to previous similar incidents, and suggesting teammates to contact for assisting users by identifying the root cause and providing resolution steps with appropriate references. Furthermore, the framework facilitates additional automation tasks, including the generation of post- incident reports that encompass all key incident timelines and the creation of Jira issues based on the identified next actions. In this paper, we will provide a comprehensive overview of the solutions developed for these recommendations and automation tasks, as well as the online and offline monitoring and evaluation steps taken to track quality and user engagement.

Ready to formalize these skills? Explore the ADaSci Certified Agentic AI System Architect (CAASA) — a proctored agentic AI certification covering orchestration, memory, evaluation, and governance for production agent systems.