Claude and Cursor Can’t Do Data Right
Table of Contents
AI coding agents are writing SQL, building pipelines, and modifying schemas at a pace that would have been unimaginable two years ago. They’re fast, tireless, and increasingly capable. But they lack something every experienced data engineer has: the operational instincts to do data work responsibly. They don’t feel accountable like humans do, and they certainly don’t stick around to ensure what they built works well.
A senior data engineer doesn’t just write correct SQL — they check whether the source table is healthy before querying it, investigate alerts when something breaks upstream, set up monitoring on new pipelines, and assess downstream impact before merging schema changes. They’ve built these instincts over years of production incidents. They’re immersed in their team’s daily operations. Agents have none of this context.
The result: an agent will confidently query a table that hasn’t been updated in two weeks. It will ignore an active incident and build a transformation on top of bad data. It will push a schema change that silently breaks a dozen downstream dashboards. It doesn’t triage, monitor, or validate — it doesn’t know to.
The problem gets worse with agents that serve business users directly. Conversational BI tools — Snowflake Intelligence, Databricks Genie, and others — use semantic and context layers to answer complex analytical questions. They’ll give a confident, well-formatted answer to “what was our churn rate last quarter?” while the underlying table has been stale for a year. Business users are even more susceptible to trusting these unvetted results than engineers are.
Introducing Monte Carlo’s Agent Toolkit
Today we’re announcing the Monte Carlo Agent Toolkit — a collection of skills and workflows that teach AI coding agents to do data right. Whether it’s validating data health, alert triage, troubleshooting, monitoring, and safe change management – it teaches your agents to do it.
The toolkit works alongside Monte Carlo’s MCP Server (25+ tools and growing quickly) and currently supports Claude Code, Cursor, OpenCode, Copilot CLI, and Codex. Skills are the expertise layer on top of the connectivity layer: while the MCP Server gives agents access to Monte Carlo, skills teach agents when and how to use it.
The toolkit covers the full lifecycle of working with data — organized around five jobs your agent needs to do:
- Trust. Before an agent touches a table, it should know if the data is healthy. The Asset Health skill gives agents a structured trust report: current status (healthy, degraded, or unhealthy), active alerts, monitoring coverage, and upstream dependency health. It’s the reflex that experienced engineers have — now built into the agent.
- Incident Response. When alerts fire, the toolkit chains triage, root cause analysis, and remediation into a single workflow. The agent scores alerts by confidence and impact, walks lineage upstream and proposes fixes with rollback plans — all without leaving your editor.
- Monitoring. Agents can analyze coverage gaps across your warehouses, recommend monitors, and create them as monitors-as-code YAML — validated against real table schemas before generation. No more invalid fields or nonexistent tables.
- Prevent. The moment you edit a dbt model, MC Prevent activates automatically via hooks: it surfaces downstream blast radius, active alerts, and a risk tier before your agent writes a line. Validation notebooks catch regressions before merge. It helps make pipeline changes confidently and fire drill free.
- Optimize. Surface your slowest queries and most wasteful assets so agents can help optimize the warehouse alongside the pipelines they’re building.
Getting started takes two commands in Claude Code:
/plugin marketplace add monte-carlo-data/mc-agent-toolkit
/plugin install mc-agent-toolkit@mc-marketplaceWe recommend enabling auto-updates so you always have the latest skills as we ship them. In Claude Code, run /plugin, select the monte-carlo-data/mc-agent-toolkit marketplace, and choose Enable auto-update. Claude Code will then refresh the toolkit automatically at the start of each session.
For Cursor and other editors, see the quickstart guide.
See It in Action
Try something like:
Triage my alerts from the last 24 hours, pleaseAnd your AI agent will quickly output something like the below, allowing you to quickly understand what needs your attention right now, what can be looked at later and what can be tuned out.
Why This Matters Now
As organizations adopt AI agents in producing and consuming data — agents building pipelines, agents answering business questions, agents operating infrastructure — the blast radius of bad data multiplies. Every agent that touches data without the proper context and skills is a potential source of silent failures.
Monte Carlo has always been about helping teams trust their data + AI. Now that extends to the agents those teams are deploying. Dozens of companies and hundreds of individuals have already connected AI agents to Monte Carlo via MCP — and we’re just getting started.
The Agent Toolkit is open source and available today:
Give it a try, star the repo, and let us know what you think.
For a deeper dive into every skill and how they compose, check out Mor Ofir’s detailed walkthrough.
Our promise: we will show you the product.
