Genie Code

Genie Code — Databricks’ agentic coding agent — writes notebooks, explores Unity Catalog, builds dashboards, and deploys apps from plain-English instructions. It extends with custom Agent Skills, and Nimble ships one: install it, describe a goal, and Genie Code turns it into a working data product — discover the right Web Search Agents, ingest live web data into a Delta table, and build an AI/BI dashboard or a deployed app. End to end, without leaving Databricks. Where Genie One answers live-web questions, Genie Code builds — it stands up the table, the dashboard, and the app for you.

​

What you get

• A brief becomes a data product. “Pricing comparison on dog products from Amazon and Walmart” → a governed Delta table, an AI/BI dashboard, and optionally a deployed app.
• Native execution, nothing extra to install. The skill runs as inline SQL over the Nimble Unity Catalog table functions, and hands dashboards and apps to Genie Code’s own dashboard agent and AppsAgent. No external tooling.
• Discovery, not guesswork. It reads the live agent catalog at runtime (nimble_agent_list, nimble_agent_describe) and probes one call per source before fanning out — so it uses each agent’s real parameters and fields.
• Set-based ingestion. A control table plus one correlated LATERAL nimble_agent_run INSERT loads every source in parallel and lands a unified, normalized Delta table — reproducible and expandable.
• The “never leave Databricks” story, end to end. From a sentence to a published dashboard, all inside Genie Code.

​

Prerequisites

• The Nimble × Databricks integration installed — the nimble_integration.tools.* table functions. See Data Enrichment for setup. The skill’s first step verifies they exist and stops with guidance if not.
• Genie Code in Agent mode — custom skills load only in Agent mode.

​

Install the skill

Genie Code loads skills from a .assistant/skills/ directory. Copy the Nimble skill folder there — keep the folder name nimble-data-products.

• User skill (just you; the prototype path): /Users/{username}/.assistant/skills/
• Workspace skill (org-wide; admins): Workspace/.assistant/skills/

The quickest path: open the skills folder from the Genie Code panel (⚙ Settings → Open skills folder) and copy the folder in, then start a new Agent-mode chat. To keep it updated, back the skills folder with a Databricks Git folder pointed at the cookbook.

​

Ask in plain English

In a Genie Code Agent-mode chat, describe the goal — the skill auto-loads by its description (or @-mention it):

• pricing comparison on dog products from Amazon and Walmart
• scrape Zillow listings for Austin into a Delta table and build a dashboard
• show competitor prices from the web in a dashboard

Genie Code confirms the deliverable (table / dashboard / app), the sources, and the target schema, then runs the flow and hands back the links and the headline insight.

​

How it works

1. Discover — nimble_agent_list() finds the agents matching your sources; nimble_agent_describe() reads each one’s exact input parameters.
2. Probe — one nimble_agent_run call per source confirms the real output fields and localization before fanning out.
3. Ingest — a control table drives a single correlated LATERAL nimble_agent_run INSERT that loads every source in parallel into one unified, normalized Delta table.
4. Build — Genie Code’s native dashboard agent assembles an AI/BI dashboard from the table; its AppsAgent can deploy an app (Python frameworks by default — Streamlit / Dash / Gradio — with React supported build-less).
5. Deliver — the published dashboard and running app links.

​

Resources