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
• Asks about investigating agent traces or conversations
• Says things like "monitor my agent", "track agent latency", "alert on agent errors"
• 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

---

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.

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.