Databricks Genie Agents skill

Databricks Genie Agents

A Genie Agent is a curated agent scoped to one business area: its tables / Metric Views, documents (UC volumes), sample questions, and instructions. It answers by writing SQL on a warehouse and, in Agent mode, by searching its documents, so one turn can combine SQL results and cited files. Distinct from Genie One (databricks-data-discovery), which answers across your data without a curated Agent.

Replaces Knowledge Assistant and Supervisor Agent for docs + data. Both are legacy ("For new low-code agents over your documents and data, use Genie Agents"). Document Q&A (was a KA) = an Agent with a volume: no endpoint, no sync. Routing between data and docs (was a Supervisor over Genie + KA) = one Agent with tables + a volume. See Documents. An Agent can also attach Databricks-provided MCP connectors (Google Drive, Microsoft 365, Slack, Jira/Confluence, GitHub, ...; Beta, set up in the UI, Agent mode only). Custom MCP servers, UC functions, or other endpoints aren't supported; build a code agent (databricks-ml-training). To maintain an existing KA / Supervisor: legacy-ka-supervisor.md.

Genie Agent Lifecycle

Phase	Reference	Load when	Typical CLI
Create / update	create-genie-agent.md	Building or editing an Agent: the workflow, the serialized_space payload format, and the CLI	discover-schema → create-space / update-space
Attach documents	create-genie-agent.md § Documents	Document Q&A, docs + data, replacing a KA / Supervisor	volumes create → update-space → probe in Agent mode
Query / converse / validate	query-genie-agent.md	Conversation API (tables), Agent mode API (needed for documents), Metric View SQL	start-conversation / Agent mode responses
Improve (inspect / diagnose / benchmark / fix)	improve-genie-agent.md	Working with an existing Agent: wrong/empty answers, quality tuning, benchmark evaluation	get-space; genie-create-eval-run; update-space
Export / migrate	genie-agent-cicd.md	Export, import, cross-workspace migration, batch migration, DABs/CI-CD	get-space → remap → create-space
Optimization history	uc-persistence.md	Setting up UC Delta tables for multi-pass optimization history (CREATE TABLE DDL only)
Legacy	legacy-ka-supervisor.md	Maintaining or migrating an existing Knowledge Assistant / Supervisor Agent	knowledge-assistants / supervisor-agents

Typical flow: create → query/validate → improve (diagnose / benchmark / fix, as needed).

Commands

An Agent is defined by a serialized_space JSON object (sources, sample questions, per-column context, instructions), kept in a local genie_agent.json and passed to the CLI. Core commands below; the lifecycle table points to the full workflow, payload schema, and flags.

databricks genie list-spaces                                          # find Agents
databricks genie get-space SPACE_ID --include-serialized-space        # inspect config
databricks genie create-space WAREHOUSE_ID "$(jq -c . genie_agent.json)" \
  --title "..." --parent-path /Workspace/Users/you@company.com/genie_spaces   # create
databricks genie update-space SPACE_ID --serialized-space "$(jq -c . genie_agent.json)"  # full-replace edit
databricks genie start-conversation --no-wait SPACE_ID "a question"   # converse / validate (tables only)
databricks genie trash-space SPACE_ID                                 # delete

Documents (replaces KA and Supervisor)

Attach a managed UC volume in data_sources.volumes[]: create-space / update-space start indexing automatically, and documents are read in Agent mode only (the chat-mode start-conversation ignores them). Setup, waiting for indexing, adding documents to an existing Agent, permissions, limits, and evaluation: create-genie-agent.md § Documents.

Prerequisites

• Tables / Metric Views in Unity Catalog and/or a managed UC volume of documents
• SQL Warehouse to execute queries (auto-detected if not specified)

Related Skills

• databricks-data-discovery: general natural-language data exploration / "ask Genie" (Genie One) across your data, when you are not targeting a specific curated Genie Agent
• databricks-metric-views: governed business metrics that Genie consumes. See its SKILL.md for design rules that affect Genie answer quality, and query-patterns.md for the MEASURE() query rules Genie follows.
• databricks-unstructured-pdf-generation: generate demo / eval PDFs to attach to an Agent as a volume
• databricks-ml-training: code agents (ResponsesAgent) when you need orchestration Genie can't express
• databricks-unity-catalog: manage the catalogs, schemas, tables, and volumes Genie uses