AgentSentinel: A Multi-Agent Framework for Autonomous Root Cause Analysis and Self-Healing in Distributed Systems

Failures in distributed systems propagate rapidly across interdependent components, yet widely adopted observability platforms such as Prometheus and Grafana provide only detection and alerting capabilities, leaving remediation entirely to human operators. This paper presents AgentSentinel, an autonomous multi-agent framework that detects anomalies, determines their root cause using large language models (LLMs), and executes targeted repairs on live infrastructure while maintaining fallback escalation paths. The framework comprises three specialized agents. The Observer Agent monitors service logs and metrics using a weighted multi-signal scoring model. The Reasoning Agent employs a ReAct-based [2] loop with an LLM and a service dependency graph to isolate the true root cause from downstream symptoms. The Repair Agent validates proposed actions against an allowlist of reversible operations and executes them on the running system. A feedback mechanism termed Reinforcement-Informed Learning from Memory (RILFM) persists resolved incident records and surfaces them as few-shot examples for future reasoning cycles. The framework was evaluated in a controlled, single-run feasibility study on a Docker Compose testbed across five fault scenarios: memory leak, service crash, CPU spike, database connection failure, and network timeout. All five faults were detected and remediated in a single ReAct iteration, with a mean time to detect (MTTD) of 1.98 seconds, a mean time to repair (MTTR) of 5.58 seconds, and zero false positives during fault-free monitoring. These results establish architectural feasibility in a controlled testbed rather than production-scale readiness.