PRISM: Pipeline for Root-cause Investigation via Specialized Multi-agents

Root Cause Analysis (RCA) in microservice architectures is a complex task requiring the correlation of multimodal telemetry data across distributed components. Existing single-agent LLM approaches often struggle with long-context reasoning and hallucination, lacking rigorous evidence verification. To address these limitations, we propose PRISM (Pipeline for Root-cause Investigation via Specialized Multi-agents), a hierarchical multi-agent framework designed to emulate the workflow of Site Reliability Engineers (SREs). PRISM decomposes the RCA process into three specialized roles: a Triage Agent that rapidly localizes suspect services using rule-based traffic analysis; an Investigator Agent that gathers objective, quantified evidence across temporal windows without making diagnoses; and an RCA Specialist Agent that synthesizes evidence to deduce the root cause. Additionally, we introduce an Evidence Search Agent that works in tandem with PRISM to iteratively refine tool configurations by collecting ground-truth evidence, ensuring high-fidelity evidence retrieval. We evaluate PRISM on a high-fidelity dataset collected from the Google Online Boutique system, covering 16 scenarios across four fault categories. Experimental results demonstrate that PRISM achieves 93.8% accuracy in service localization and 100% in fault type classification, validating the effectiveness of structured multi-agent collaboration in automated troubleshooting.