Building Traffic AI — Part 2: From Anomaly Detection to Proactive Platform Operations

In Part 1 , we described how Traffic AI evolved from Q&A automation into a system capable of using organizational knowledge and operational tools through MCPs. By that point, the agent could investigate problems across Grafana, ClickHouse, Centralized Load Balancing Management API and anomaly signals. In Part 2, we focus on what happened when we started making those capabilities proactive and production-ready: automatically triggered anomaly investigations, the architecture behind long-running AI workloads, the evolution of our knowledge base, and the engineering lessons we learned along the w