Agentic Data Pipeline Monitoring: A Multi-Agent Architecture for Anomaly Detection in Financial Transaction Systems

Data pipeline reliability is a first-order concern in financial transaction systems, where silent failures — pipelines that complete successfully while delivering degraded or corrupted data — can propagate undetected for hours before surfacing in downstream business processes. Traditional threshold-based monitoring cannot detect failures that were not anticipated at configuration time, leaving financial institutions exposed to the most dangerous failure modes: the ones nobody thought to look for. This paper presents a production-verified multi-agent architecture for financial data pipeline monitoring. The system decomposes the monitoring problem across four specialized agents: a Baseline Agent that learns normal behavioral signatures for each pipeline using rolling statistical models; a Detector Agent that identifies deviations from learned baselines using ensemble anomaly scoring; an Explainer Agent that generates natural language descriptions of detected anomalies with explicit citations to the underlying metrics; and an Orchestrator Agent that coordinates escalation, routes alerts, and manages the human-in-the-loop checkpoints required for financial data governance. Deployed in production across enterprise financial data pipelines processing multi-terabyte transaction workloads at a major US financial technology company, the system achieved a 40% reduction in incident response time compared to the threshold-based system it replaced. More significantly, the architecture detected anomalies that the prior system could not have detected — novel failure modes that emerged in production and were not represented in the threshold configuration. We describe the architecture, the key engineering decisions that make it safe for financial production environments, and the evidence that multi-agent decomposition outperforms both monolithic LLM approaches and traditional statistical monitoring for this problem class. We discuss generalization patterns for other critical infrastructure domains including healthcare, energy grid telemetry, and supply chain systems. The reference implementation is released as open source under MIT license as Jasus-Pulse (https://github.com/suhasgr09/jasus-pulse) to enable reproducibility and adoption by the broader data engineering community.