RADAR started as a response to ticket backlogs that spanned traffic, retrieval, pacing, ranking, and bidding. Each domain had its own signals and observability quirks, so we trained specialized agents on 1K–2K labeled tickets per area and exposed them to the right logs, metrics, and configs.

A coordinator agent aggregated their findings, reasoned over conflicts, and produced a single diagnosis plus mitigation steps. The system reduced ticket resolution time by roughly 20% and gave vendors a consistent, auditable workflow during peak events.

Two lessons: (1) multi-agent systems shine when inputs are heterogeneous and domain-specific; (2) invest early in tooling to keep agents aligned with data contracts—drift in observability pipelines can quietly erode quality.