Genie Code for data science: Work at the speed of your ideas

Early customer feedback reports moving from raw data to governed tables, tracked models, and a production-ready serving endpoint in about 90 minutes, with productivity gains of three to ten times on common tasks. The agent is grounded in Unity Catalog and fluent across data science, data engineering, machine learning, and BI.

Workspace-aware automation that does more than write code

Genie Code uses agentic search across Unity Catalog metadata, sample data, and existing notebooks, dashboards, and queries to select relevant assets and explain why they fit the task. It drafts plans, edits notebooks, creates visualizations, and auto-logs training runs to MLflow. When execution fails, the agent diagnoses errors, adjusts code, and retries without handholding.

Approval controls match different risk postures. Users can approve step by step, allow tools per thread, or enable auto-approve so the agent evaluates each tool call against the prompt's intent, reducing click fatigue while preserving oversight.

Teaching the agent your conventions: instructions, skills, and connectors

Teams can shape Genie Code into a tenured coworker by codifying how they work.

• Instructions are markdown files loaded into every conversation, capturing workspace standards like visualization palettes, pipeline preferences, and other best practices. Admins can set workspace instructions so guidance applies by default to all users.
• Skills are markdown files loaded only when their description matches the task, such as a model card template for ML workloads. Selective loading keeps context focused and avoids irrelevant guidance.
• Connectors bring in external context and capabilities. Built-in OAuth covers Gmail, Calendar, Drive, Slack, and Atlassian. MCP servers can be added, including Unity Catalog Functions, Databricks AI Search, and Genie Agents. External MCP servers can be made available through a Unity Catalog connection.

 

A practical loop emerges: start a task in the agent, correct it as it works, then ask Genie Code to save the successful approach as a reusable skill so future runs execute more autonomously.

Scaling up from a side panel to a command center

Genie Code now runs as a full-page command center with multiple parallel threads. One thread can profile a table while another builds visualizations and a third converts a Jira ticket into an AI/BI dashboard. The agent can also create Databricks Lakeflow jobs to schedule recurring analyses and integrates with MLflow to compare runs, surface insights, and recommend next steps such as hyperparameter tuning.

Automations are coming soon. Users will be able to schedule Genie Code to work through open tickets or recurring tasks overnight, then review results in the morning. Combined with Genie ZeroOps for maintaining pipelines and models, the long-term direction is clear: agents handle the mechanical execution so humans focus on judgment, prioritization, and interpretation.

Turning context into leverage, not overhead

Genie Code compounds in value as teams teach it their norms and connect it to their systems. Start more tasks in the agent, guide it when needed, and promote effective patterns into instructions and skills. Investment in context shifts the human role from clicking through steps to deciding what matters next.