OPS CORTEX architecture separates root-cause detection from LLM explanation

Modern microservice failures are easy to detect but hard to explain—alerts flood, dashboards multiply, and engineers waste time reconstructing context. OPS CORTEX, a working multi-agent prototype by Momil Seedat, flips this dynamic by introducing operational memory: a persistent, structured representation of normal system behavior, service dependencies, and historical failures. The system divides root-cause analysis into two distinct tasks: deterministic derivation from a learned dependency graph and temporal threshold crossings, followed by an LLM that only explains, confirms, and recommends using pre-assembled evidence. This separation avoids confounding the two tasks, a common pitfall in existing tools.

OPS CORTEX organizes operational memory into four tiers and is validated on an instrumented e-commerce benchmark with eight injectable failure scenarios, mapping each architectural choice to documented production cascading failures. The paper reviews literature on observability, anomaly detection, graph-based localization, and LLM-assisted diagnosis, showing how each design decision mitigates specific real-world failure modes. By keeping the root-cause computation deterministic and the LLM role confined to explanation, OPS CORTEX reduces noise and engineer overhead, promising a more reliable and scalable approach to autonomous operations.