Snowflake Cortex Agents emit rich trace data via the SNOWFLAKE.LOCAL.GET_AI_OBSERVABILITY_EVENTS table function, covering span types like chat, planning, response generation, and tool calls. Key performance metrics to monitor include token consumption, duration, and status codes at the span level. Common failure patterns include token spikes from multi-turn context accumulation or retrieval config changes, and usage volatility that can mask elevated error rates. Combining signals across span types — e.g., high planning tokens alongside low tool-call completion rates — enables faster root cause analysis. The post also introduces Monte Carlo's agent observability integration with Snowflake Intelligence for automated monitoring at scale.
Table of contents
How to identify and fix agent performance issues using Cortex Agent telemetryObserving your Cortex Agent fleetWhat Snowflake logs for Cortex Agents: a look under the hoodMapping records to span typesGet Michael Segner’s stories in your inboxKey performance metrics to monitorTotal tokensDurationStatus CodesCombining SignalsCommon agent performance issuesToken spikes: what they look like and what causes themUsage volatilityThe underlying principleSort: