LLM-free Jev pipeline diagnoses K8s faults in 14.6s, passing 76.2% of 105 diagnoses

bibryam · x · 2026-10-13

The SREGym team details a Jev-driven diagnosis pipeline that skips the LLM agent entirely for Kubernetes root-cause analysis.

A programmatic collector reads K8s objects, events, pod logs, and resource usage, grouping observations by component and summarizing failure signs
Jev first picks which component to investigate, then receives numbered evidence items and classifies the component as origin, downstream victim, or unrelated, selecting cause category and key evidence
Victims trigger a follow-up on the linked dependency; unrelated candidates are skipped, looping until the origin is confirmed
Across 21 SREGym-Lite faults, the pipeline passes 80 of 105 diagnoses (76.2%) with a median diagnosis time of 14.6 seconds

More from coding & agent