• Sources: primary
  • Summary: A calibration agent with a capable model maps the telemetry, mines existing alert rules and dashboards, measures normal ranges over the past week, and saves validated checks once. Each later run uses plain code to query metrics, compare against baselines and track Loki log patterns, then makes one tool-free call to a small model that decides whether to notify. Grafana says 41 of its 43 internal watchers cost under $10 a month at $2 per million tokens, and one watcher found more than 10 customer-affecting production bugs.
  • Why it matters: It is a concrete cost pattern for agentic monitoring: spend model reasoning once at setup and keep the per-run model call narrow.