Designing Multi Agent Workflows for Root Cause Analysis in Distributed Systems

Root cause analysis in a distributed system rarely fails because telemetry is entirely absent. More often, the evidence is fragmented across metrics, logs, traces, deployment records, and service ownership boundaries. A single general-purpose agent can search these sources, but it tends to mix observation with interpretation and may settle on the first plausible explanation. This paper proposes a multi-agent workflow that separates evidence collection, hypothesis formation, adversarial verification, and incident reporting. Specialized agents operate over a shared incident record and must attach provenance to every material claim. A coordinator controls scope and budget but cannot declare a root cause by itself. A verifier tests competing hypotheses against topology, time, and counterevidence before a human incident commander accepts the diagnosis. The design treats the language model as a reasoning aid rather than an authority. It is intended as a reference architecture; no production deployment or comparative experiment is claimed.

Artificial Intelligence Cloud Computing Software Engineering