Analyze coverage gaps and deploy data quality monitors

A Monte Carlo skill that handles all monitoring requests - coverage analysis, data monitor creation, and AI agent monitoring - by routing to the right reference file based on user intent. It calls Monte Carlo exclusively through the plugin's bundled MCP server with fully-qualified tool names, never a separately-configured monte-carlo-mcp server. Across its three tool categories it spans 26 tools: 14 coverage and discovery tools (get_warehouses, get_use_cases, get_use_case_table_summary, get_use_case_tables, get_monitors, get_asset_lineage, get_audiences, get_unmonitored_tables_with_anomalies, search, get_table, get_queries_for_table, get_field_metric_definitions, get_domains, get_validation_predicates), 5 data monitor creation tools that each follow a two-call preview-then-confirm pattern (dry_run=True renders MaC YAML, dry_run=False deploys and returns a deep link), and 7 agent monitoring tools covering agent metadata, conversation and trace inspection, plus four agent monitor-creation tools for metric, evaluation, trajectory, and validation monitors. The primary coverage workflow discovers warehouses, then either explores Monte Carlo's auto-generated use cases (sorted HIGH/MEDIUM/LOW by criticality, defaulting scope to HIGH+MEDIUM without asking) or falls back to importance-based table discovery when no use cases exist, cross-referencing get_unmonitored_tables_with_anomalies to prioritize tables with real detected issues over inactive ones. When coverage analysis leads to monitor creation, it dedups against existing tag-scoped monitors first (the backend upserts on description+domain, so a same-description monitor silently overwrites), asks only one question (which notification audiences to use), and defaults new monitors to draft rather than active. It also recognizes transient truncate-and-reload tables (0 rows on direct query despite recent Monte Carlo write activity) to avoid wasted investigation, and degrades gracefully - explaining and offering a next path rather than stopping - when a tool returns empty or a prerequisite like a database MCP or use-case data is missing.

Use it for monitoring coverage or coverage-gap questions, creating any data or agent monitor, exploring use cases, or setting up alerts on agent latency, quality, or execution behavior. Do NOT use it for querying or exploring table contents, triaging or responding to active alerts (the prevent skill's Workflow 3 handles that), running pre-change impact assessments (prevent skill's Workflow 4), or inspecting/editing existing monitor configuration directly via get_monitors.

Input is a monitoring request - a coverage question, a specific table/field/metric to monitor, or an agent to monitor - plus, for creation flows, the user's choice of notification audience(s). Output is either a coverage analysis (criticality breakdown, monitored-vs-unmonitored counts, prioritized gap list) or a monitor preview/deployment: rendered YAML on dry-run, then a live monitor with a deep link once confirmed.

Runs entirely through Monte Carlo's bundled MCP server (monte-carlo-mcp is required and must be authenticated); an optional database MCP server (Snowflake, BigQuery, Redshift, or Databricks) enables SQL query-pattern profiling of candidate tables, and is skipped gracefully - without asking the user to configure one - if unavailable.

Data teams who need to find monitoring coverage gaps, prioritize which tables or AI agents to monitor first based on criticality and real anomaly activity, and deploy Monte Carlo monitors through a guided, dedup-aware, audience-routed creation flow rather than authoring MaC YAML by hand.