- Published on
Monitoring, Troubleshooting, and Performance Patterns
- Authors
- Name
- DeepHorus
Snowflake's CoWork AI assistant, designed to simplify data interactions, introduces a significant monitoring challenge: usage data is fragmented across four separate Account Usage views, each with a 45-minute latency. According to a new technical guide from Snowflake Chronicles, this fragmentation can mislead administrators on cost and system health if they rely on a single view. The guide, part of a five-article series on CoWork deployment, addresses the critical gap between building and operating such systems.
CoWork allows users to query data via natural language, but each interaction triggers multiple billable components: a language model call, Cortex Search retrieval, SQL generation on a warehouse, and guardrails scanning. Each of these components logs data in a different Snowflake Account Usage view, including CORTEX_AGENT_USAGE_HISTORY, SNOWFLAKE_INTELLIGENCE_USAGE_HISTORY, CORTEX_AI_FUNCTIONS_USAGE_HISTORY, and CORTEX_AI_GUARDRAILS_USAGE_HISTORY. The key insight is that to trace a single CoWork question to its agent execution, one must join SNOWFLAKE_INTELLIGENCE_USAGE_HISTORY with CORTEX_AGENT_USAGE_HISTORY using the REQUEST_ID and PARENT_REQUEST_ID columns.
The guide emphasizes that failing to combine these views leads to incomplete cost tracking and missed health signals. For example, CoWork-initiated requests appear in SNOWFLAKE_INTELLIGENCE_USAGE_HISTORY, while programmatic agent calls appear in CORTEX_AGENT_USAGE_HISTORY. Monitoring only one view could result in underestimating credit consumption by half. Additionally, the 45-minute latency in Account Usage views makes them unsuitable for sub-minute alerting, though they are adequate for daily operations and trend detection.
To address these challenges, the guide proposes a dedicated monitoring infrastructure. It recommends creating a separate XSMALL warehouse (cowork_monitoring_wh) to ensure observability queries never compete with agent workloads for compute. This warehouse costs roughly 0.03 credits per minute when active and auto-suspends after 60 seconds, minimizing costs. The guide also includes SQL code for building health dashboards and automated alerts using CORTEX_AGENT_USAGE_HISTORY as the primary telemetry view.
Industry experts note that Snowflake's documentation does not fully address this monitoring complexity, leaving users to piece together solutions. The guide, validated against a live Snowflake account as of June 2026, fills a critical gap. It also covers proactive detection of runaway AI queriesâthose still running but already consuming significant creditsâusing the IS_COMPLETED column in scheduled tasks.
Looking ahead, Snowflake's Cortex AI Guardrails, released in May 2026 for Enterprise Edition, adds prompt injection detection with minimal cost and configuration. However, real-time sub-minute alerting remains out of scope and would require event tables. The guide concludes that while monitoring CoWork is challenging, a structured approach using dedicated warehouses and joined views enables effective oversight, paving the way for more robust production deployments.
