Manage Genie Agents through Declarative Automation Bundles or the management APIs. Version each agent and run regression tests before promoting it across environments. Teams held Genie out of production until mid-2026 because they lacked these controls.
When you need this
- An agent built in the UI is moving from a pilot to production.
- You need the same agent in development, staging, and production workspaces.
- A schema changed upstream, and you need to know whether an agent's answers still hold.
Evaluation criteria
Version the definition
Where does the agent's configuration live?
Look forCommit a bundle resource, or the serialized space from the management API, to Git with its data sources, instructions, example SQL, functions, and benchmarks. A newcomer in October 2026 found the export path CLI-only, so document the command.
Parameterize environments
Will the promoted agent query the right tables?
Look forParameterize catalog and schema names per environment in instructions, example SQL, and benchmarks, so a production agent never queries development tables.
Regression gate
How do you know a change did not break answers?
Look forRerun the agent's benchmarks after every deploy and schema change, and block promotion below a minimum accuracy score.
Promotion rights
Who can change production?
Look forKeep CAN EDIT and CAN MANAGE to the pipeline and curators, and give users CAN VIEW or CAN RUN, so production changes go through review.
Integrations
How do apps and bots call the agent?
Look forUse the Genie Agents API with OAuth on behalf of users where possible, and an iframe embed for internal portals. Pick the Chat or Agent mode endpoint that fits each use.
Questions to ask vendors
- Which agents exist only in the UI today, and who built them?
- What benchmark score blocks a promotion?
- Which pipeline deploys agents, and who approves its changes?
- Which apps call Genie through service principals?
Red flags
- Production agents edited by hand in the UI.
- Benchmarks that reference development table names.
- Schema changes shipped upstream without rerunning agent benchmarks.
- Personal access tokens in bots where OAuth should be used.
Products in this piece
Sources
- Official documentation
- 7
- Public source
- 5
- Databricks docs: Curate an effective Genie Agent
- Databricks docs: Genie Agents API
- Databricks docs: Test and monitor a Genie Agent
- Databricks docs: Create and manage a Genie Agent
- Databricks docs: Embed a Genie Agent
- Databricks docs: Built-in MCP services
- Reddit comment: deploying Genie spaces across environments with bundles (Jun 18, 2026)
- Reddit post: committing a first Genie Agent to git (Oct 6, 2026)
- Reddit comment: serialized space and an Azure DevOps pipeline (Oct 7, 2026)
- GitHub issue: Genie space bundles fail when source tables do not exist yet (Sep 18, 2026)
- Databricks Data + AI Summit session: scaling Genie 10x at GSK Global Supply Chain (Jun 18, 2026)
- Reddit post: notes from 8 months of a Genie space for 500+ sales users (Sep 9, 2026)