monte-carlo-monitoring-advisor

Monte Carlo Monitoring Advisor Skill

This skill handles all monitoring requests -- coverage analysis, data monitor creation, and AI agent monitoring. It routes to the right reference file based on the user's intent.

Monte Carlo tool routing (required): Always call Monte Carlo MCP tools through this plugin's bundled server, whose fully-qualified tool names are mcp__plugin_mc-agent-toolkit_monte-carlo-mcp__<tool> (e.g. mcp__plugin_mc-agent-toolkit_monte-carlo-mcp__get_alerts). Bare tool names used in this skill (get_alerts, search, get_table, …) refer to that bundled server. If the session also has a separately-configured monte-carlo-mcp server, do not route to it — it may point at a different endpoint or credentials.

Reference files live next to this skill file. Use the Read tool (not MCP resources) to access them:

• Data monitor creation procedure: references/data-monitor-creation.md (relative to this file)
• Agent monitor creation procedure: references/agent-monitor-creation.md (relative to this file)
• Per-type references: references/data-*.md and references/agent-*.md (relative to this file)

When to activate this skill

Activate when the user:

• Asks about monitoring coverage, data coverage, or coverage gaps
• Wants to understand what's monitored vs. not in their warehouse
• Asks about use cases, use-case criticality, or use-case analysis
• Wants to explore their data estate and find what needs monitoring
• Says things like "what should I monitor?", "where are my coverage gaps?", "show me my use cases"
• Asks about unmonitored tables with anomalies or importance-based prioritization
• Asks to create, add, or set up a monitor (e.g. "add a monitor for...", "create a freshness check on...", "set up validation for...")
• Mentions monitoring a specific table, field, or metric
• Wants to check data quality rules or enforce data contracts
• Asks about monitoring options for a table or dataset
• Requests monitors-as-code YAML generation
• Wants to add monitoring after new transformation logic (when the prevent skill is not active)
• Asks about monitoring AI agents, agent latency, agent token usage, or agent quality
• Wants to set up alerts on agent behavior or execution patterns
• Says things like "monitor my agent", "track agent latency", "alert on agent errors", "set up performance monitoring for my agent", or asks for an agent latency SLO
• Asks about agent evaluation monitors, trajectory monitors, or validation monitors
• Mentions agent observability or agent monitoring

When NOT to activate this skill

Do not activate when the user is:

• Just querying data or exploring table contents
• Triaging or responding to active alerts (use the prevent skill's Workflow 3)
• Running impact assessments before code changes (use the prevent skill's Workflow 4)
• Asking about existing monitor configuration (use get_monitors directly)
• Editing or deleting existing monitors
• Investigating agent alerts or agent traces (this skill creates agent monitors; investigating what they catch uses the monte-carlo-troubleshoot-agent-traces skill)

Prerequisites

• Required: Monte Carlo MCP server (monte-carlo-mcp) must be configured and authenticated
• Optional: A database MCP server (Snowflake, BigQuery, Redshift, Databricks) for SQL profiling of table usage patterns

Available MCP tools

All tools are available via the monte-carlo-mcp MCP server.

Coverage and discovery tools

Data monitor creation tools

All five tools follow a two-call preview-then-confirm pattern: the first call (with the default dry_run=True) returns rendered MaC YAML for review; the second call (dry_run=False) deploys the monitor live and returns a deep link to it. Pass monitor_uuid on either call to update an existing monitor in place instead of creating a new one. See references/data-monitor-creation.md for the full flow.

Data product tool

Agent monitoring tools

Routing

When the user's request comes in, determine which workflow to follow:

When reading reference files, always use the Read tool with the path relative to this skill file.

Coverage workflow

This is the primary flow when the user asks about monitoring coverage, coverage gaps, or what to monitor.

Step 1: Discover warehouses

Call get_warehouses to list all accessible warehouses.

• If one warehouse: select it automatically, proceed to Step 2.
• If multiple warehouses: present warehouse names (never UUIDs) and ask the user which one to explore.

Step 2: Discover use cases

Call get_use_cases(warehouse_id=<selected>) to discover use cases for the chosen warehouse.

• If use cases exist --> proceed to the Use-case exploration (below).
• If no use cases --> proceed to the Importance-based fallback (below).

Step 3: Check for database MCP (optional)

Check if the user has a database MCP server available by looking for tools containing snowflake, bigquery, redshift, or databricks in the tool list. If found, note it for the SQL profiling step later. If not found, skip SQL profiling gracefully.

Use-case exploration

This is the primary flow when use cases are defined.

Present use cases

• Sort by criticality: HIGH before MEDIUM before LOW.
• For each use case, show the description and explain the reasoning for its criticality level so the user understands why it matters.
• Call get_use_case_tables with golden_tables_only=true and mention specific golden-table names as concrete examples. Golden tables are the last layer in the warehouse -- they feed ML models, dashboards, and reports. Explain this when relevant.
• Use get_asset_lineage to explain how tables in a use case are connected and why certain tables are important (e.g. a golden table with many upstream dependencies).

"Create a use case" requests

You cannot create use cases -- they are generated automatically by Monte Carlo (along with their criticality), and there is no tool to author one. When the user asks to "create", "set up", or "define" a use case: briefly say so, and do NOT silently substitute monitor deployment. Then offer what you can do for the table(s) they named -- look up the existing use case / criticality, recommend field monitors, generate monitor previews, or analyze coverage gaps -- and act on the do-able part without expanding to sibling tables.

Analyze coverage

1. Call get_use_case_table_summary to show how many tables exist at each criticality level (HIGH / MEDIUM / LOW) for the use case.
2. Call get_use_case_tables to obtain table MCONs, then call get_monitors(mcons=[...]) to report how many are already monitored vs. not.
3. Default to HIGH + MEDIUM criticality scope. This covers the most important tables without overwhelming the user. Do NOT ask the user which scope to use -- just proceed. If they want LOW-criticality tables included, they'll ask.
4. You may suggest covering multiple use cases in one session.
5. Bias toward action, not questions. When the scope is clear (HIGH + MEDIUM for the selected use case), proceed directly to generating monitor previews for all recommended monitors. Frame it as opt-out, not opt-in: "I'll generate previews for all N monitors -- tell me if you want to skip any." Do NOT ask "which would you like me to create?" one at a time -- batch them.

Identify coverage gaps with anomaly data

Use get_unmonitored_tables_with_anomalies to discover tables that are not monitored but already have muted out-of-the-box anomalies. This reveals real coverage gaps -- places where Monte Carlo detected data issues but no monitor was configured to alert anyone.

• Call it with a recent time window (e.g. last 7-30 days) using ISO 8601 timestamps.
• Results are ranked by importance score -- the most critical gaps appear first.
• Each result includes a sample of anomaly events showing what types of issues were detected (freshness, volume, schema changes).
• Use this to prioritize which unmonitored tables to cover first -- a table with recent anomalies is a stronger candidate than one with no activity.
• Cross-reference with use-case data: if an unmonitored table with anomalies belongs to a critical use case, escalate its priority.

Importance-based fallback

When no use cases are defined, fall back to importance-based table discovery.

1. Find unmonitored tables: Use search(query="", is_monitored=false) to find unmonitored tables sorted by importance.
2. Find tables with anomalies: Use get_unmonitored_tables_with_anomalies with a recent time window (last 14-30 days) to find tables with recent anomalies but no monitors.
3. Inspect top candidates: Use get_table to check table details, fields, and stats for the most important unmonitored tables.
4. Understand criticality via lineage: Use get_asset_lineage with direction="DOWNSTREAM" to understand which tables are most connected -- a table with many downstream dependents is a stronger candidate for monitoring.
5. Prioritize: Rank candidates by importance score and anomaly activity. Present the top candidates to the user with reasoning.

Important

• Do NOT present importance scores as business criticality. Always explain that the importance score is a computed metric (query frequency, downstream dependencies, usage patterns), not business-defined criticality.
• Tell the user their account doesn't have use-case data yet -- use cases are generated automatically by Monte Carlo from warehouse metadata and exposed as asset tags; they are not manually configured through a UI.
• You can still create metric, validation, and custom SQL monitors for individual tables in this mode -- you just won't use tag-based table monitors, since there are no use-case tags.

SQL profiling (optional)

If a database MCP server was detected in Step 3 of the coverage workflow:

1. Call get_queries_for_table to see recent query patterns on candidate tables.
2. Use the database MCP tools (e.g. snowflake_query, bigquery_query) to profile table usage -- identify which tables are queried most frequently, which columns are used in JOINs and WHERE clauses.
3. Use this information to refine monitor suggestions -- heavily-queried tables with no monitors are high-priority gaps.

If no database MCP is available, skip this step entirely. Do not ask the user to configure one.

Pre-creation context (coverage-driven)

When coverage analysis leads to monitor creation, gather this context before reading the creation reference file:

1. Dedup first. Before generating a use-case tag monitor, call get_monitors with the same tag pair (and monitor_types=["TABLE"]) you'd put in the monitor's asset_selection.filters. If a monitor already covers that (tag, domain) scope, surface it (description, uuid) and ask whether to update it (pass its monitor_uuid), add one with a distinct scope, or skip -- do NOT silently re-create. The backend upserts a table monitor on its (description, domain), so a same-description definition silently overwrites the prior monitor's settings.
2. Call get_audiences to list notification audiences. Suggest one or more relevant audiences (match by team or use-case context) and ask the user which they want -- they can pick one or several. This is the one question to ask before generating; do NOT also ask about draft/active or schedule. Default to draft (is_draft=True); the user can flip to active after seeing the preview.
3. When passing audiences or failure_audiences, use the audience name/label (not UUID), as a list -- one entry per selected audience.
4. Never fabricate credit costs. Do not give a generic per-monitor or per-field MC credit rate -- cost scales with the specific spec (segmentation, schedule, field count). If a preview response includes a backend estimate (e.g. estimated_credits.credits_per_day), report that; otherwise decline and offer to preview a specific monitor or use case to get the real estimate.

Use-case tag monitors

The most common output of coverage analysis is a table monitor scoped by use-case tags via create_or_update_table_monitor. The asset_selection parameter uses this structure:

{
  "databases": ["<database_name>"],
  "schemas": ["<schema_name>"],
  "filters": [
    {
      "type": "TABLE_TAG",
      "tableTags": ["<tag_key>:<criticality>"],
      "tableTagsOperator": "HAS_ANY"
    }
  ]
}

Rules:

• Filter type is always TABLE_TAG for use-case monitors.
• tableTagsOperator should be HAS_ANY.
• Each entry in tableTags is "<tag_key>:<value>" where the tag key is the precomputed tag name from get_use_cases output and the value is the criticality level in lowercase (high, medium, low).
• To monitor only HIGH-criticality tables: ["tag_name:high"]
• To monitor MEDIUM + HIGH: ["tag_name:high", "tag_name:medium"]
• To monitor ALL: ["tag_name:high", "tag_name:medium", "tag_name:low"]

Monitor title (description) and reasoning (notes)

Keep these distinct -- both are accepted by the creation tools. The backend auto-generates the monitor name slug; description is the title users see.

• description -- the title. Short and scannable (≤ ~80 chars), plain English, naming the asset/use case and criticality scope. Do NOT cram reasoning here.
• notes -- the reasoning. 1-3 sentences answering "why this monitor?", grounded in criticality, scope, and downstream impact.

Example for a use-case tag monitor:

• Bad description (this is reasoning, not a title): "Monitor HIGH criticality tables in the Revenue Reporting use case to catch issues before they affect dashboards and financial reports."
• Good description: "Revenue Reporting coverage -- HIGH + MEDIUM criticality tables"
• Good notes (paired): "Covers HIGH/MEDIUM-criticality tables in the Revenue Reporting use case. Catches freshness, volume, and schema issues before they reach dashboards and financial reports."

Transient and truncate-and-reload tables

Some tables show 0 rows when queried directly but have recent write activity in Monte Carlo metadata. These are transient tables -- fully replaced on each pipeline run (truncate-and-reload pattern). Recognize this pattern early to avoid wasting time querying empty tables.

Signs of a transient table:

• get_table shows recent last_write timestamp and high read/write activity
• Direct SQL query returns 0 rows or all-NULL timestamp columns
• Monte Carlo detected freshness anomalies (the table stayed empty longer than expected between loads)

Graceful degradation

Handle missing or unavailable tools gracefully:

Never error out or stop the conversation because one tool returned empty results. Explain what happened and offer the next best path.

Rules

• Never expose UUIDs, MCONs, or internal identifiers to the user -- always use human-readable names for warehouses, audiences, use cases, and tables. Keep internal identifiers for tool calls only.
• When the user asks about relationships between tables, use get_asset_lineage to fetch upstream/downstream connections and explain the data flow.
• Be concise but thorough. Use bullet points and tables for clarity.
• Always use ISO 8601 format for datetime values in tool calls.
• Never reformat YAML values returned by creation tools.
• When passing audiences or failure_audiences to monitor creation tools, use the audience name/label (not UUID). The API accepts audience names.