GALA: Graph-Augmented LLM Agents for Root Cause Analysis and Incident Response in Microservices

Modern microservice architectures introduce significant operational complexity, making it difficult to pinpoint the origin of failures when symptoms propagate across multiple services. Traditional root cause analysis (RCA) methods often struggle with heterogeneous telemetry data, while unconstrained LLM-based agents frequently suffer from hallucination and inefficient, non-topological exploration. Furthermore, existing systems often focus solely on ranking faulty services, failing to provide the actionable incident response and causal explanations required by site reliability engineers (SREs).

GALA+ addresses these challenges through a four-phase agentic workflow that grounds LLM reasoning in the system's actual topology.

1. Initial Hypothesis Generation: The system combines metrics-based causal inference with STRIX, a novel trace-aware scoring module. STRIX evaluates services based on three dimensions: instability (latency tail-to-median ratio), centrality (structural importance), and impact (local processing delay), using Reciprocal Rank Fusion to generate a robust initial candidate list.
2. Context Synthesis: Raw telemetry is distilled into compact, token-efficient diagnostic bundles containing JSON-formatted metrics, annotated dependency subgraphs, and deduplicated error logs.
3. Graph-Guided Investigation: Investigator agents perform depth-bounded, topology-constrained searches. By following fault-propagation edges in the dependency graph, the agents maintain focus on causally plausible paths, preventing the drift common in unconstrained LLM agents.
4. Incident Response: The framework produces structured incident summaries and stratified remediation recommendations, which are evaluated using SURE-Score, a specialized framework designed to measure RCA-specific output quality.

Evaluated on the OnlineBoutique and TrainTicket benchmarks, GALA+ significantly outperforms existing LLM-based baselines, achieving a 25 percentage point improvement in AC@1. The framework's outputs were validated through both automated dual-LLM grading and independent human SRE assessments, demonstrating superior performance in providing actionable, causally grounded insights.