Komputer.AI
ObservabilityMonitoring

Per-agent labels

Cardinality trade-off when enabling per-agent label breakdown on metrics.

By default, the agent_name label is present but empty on per-agent metrics. This keeps dashboards schema-stable — queries like sum by (agent_name) (...) work whether or not the feature is enabled, they just resolve to a single bucket when off.

To get real per-agent breakdown:

metrics:
  perAgentLabels: true

Cost: every per-agent metric series multiplies by the number of agents you create. For 100 active agents, the komputer_agent_tasks_total series count goes from ~12 (3 outcomes × 4 models) to ~1200. Prometheus handles this fine until you hit hundreds of thousands of series.

Recommendation: keep this off for production, enable in pre-prod or for one-off debugging.