The new Genie runs jobs on its own. What to set up before you let it

On 11 November 2026 Databricks published The new Genie: autonomous AI for complex enterprise work. The short version: Genie is no longer only something you ask questions. You can hand it a job, let it run on a schedule or a trigger, and approve its actions as it goes.

That is a real change in what a Genie rollout involves. Below is what was announced, what the companion posts say is generally available, and what we would put in place before switching any of it on.

What was announced

Genie as an autonomous coworker. You delegate a job through Genie One, email, Slack or Microsoft Teams. Genie can run it on a schedule or when something happens, such as a metric dipping or more than ten customers asking for the same feature. It tracks progress, chases missing inputs, and can run many jobs at once from a control panel. The post’s examples are watching insurance claims for fraud patterns and loss-ratio drift, and managing campaigns with weekly budget changes proposed for approval.

Genie Ontology. A business context layer that agents build mostly from your existing dashboards, queries and notebooks, and that keeps learning from use. It ranks governed definitions above learned ones (Databricks calls the ranking OntoRank) and is organised with Unity Catalog Domains, Pages and Ontology Snippets.

Genie Search. Hybrid search over indexed sources (Google Drive, Jira, Confluence, Workday, SharePoint) plus live search through MCP, with source permissions enforced through Unity Catalog.

Genie One additions. Artifacts (slide decks, dashboards and interactive pages that scheduled tasks keep current), Skills (repeatable workflows stored in Unity Catalog that run with the user’s permissions), and MCP connectors for Slack, Zoom, Microsoft 365, Jira, GitHub and calendars.

Genie Agents (formerly Genie Spaces) can now search PDFs and slides in Unity Catalog Volumes with page-level citations and no separate vector index. They can also be shared with partners and customers, including through Genie Clean Rooms, which limit what a recipient can see.

Genie outside Databricks. Genie plugins for ChatGPT, Claude, Copilot and Grok, and Genie MCP for connecting other AI tools directly.

Unity Gateway. One control plane for governing Genie, coding agents, models, MCP tools and skills: tool access controls, approval requirements, tracing, auditing, and per-workspace and per-user budgets with caps and alerts.

What is actually GA

The main post does not label release stages. The companion posts do, in part:

• Generally Available: Genie Ontology, Genie Search, Unity Catalog Domains and Pages, Ontology Snippets, organisational directories (starting with Workday), and Genie MCP.
• Beta: the SharePoint connector for Genie Search, and external knowledge sources for Genie Ontology.
• Coming soon, in Beta: Ontology Manager, the visual interface for curating context.
• Not stated: the Genie One post does not give a release stage for autonomous jobs, Artifacts, Skills or the Slack and Teams experience. It points to the documentation for current availability.

Check your own workspace and region before you promise any of this to a business unit.

About the 91% number

Databricks reports that in a benchmark built with Turing, Genie with Genie Ontology answered correctly in 91% of task runs. That is 1.6 times the stronger of two frontier agents on the same tasks, using less than half the tokens. The Ontology post adds an important caveat: the results reflect an established deployment where teams had already built up that knowledge.

Read it as a ceiling, not a starting point. A new workspace with undocumented tables and three competing definitions of “active customer” will not start at 91%.

What we would do before turning on autonomous jobs

A Genie that answers questions wrongly costs someone a bad afternoon. A Genie that acts on a schedule can do the wrong thing every Monday until someone notices. So the order matters.

1. Fix definitions first. Genie Ontology ranks governed definitions above learned ones, so write the governed ones down. Put key measures in metric views, and write a Unity Catalog Page for each term your business argues about. Otherwise the ontology learns from whichever dashboard is used most, which is not the same as the correct one.
2. Set approval rules in Unity Gateway before the first job. Decide which actions need a human to approve them, which MCP tools are allowed, and what each workspace and user may spend. It is much easier to loosen these later than to tighten them after an incident.
3. Start with read-and-report jobs. A scheduled job that watches a metric and writes a summary is low risk. Jobs that change budgets, open tickets or message customers should come later, with approval required.
4. Treat Skills like code. They live in Unity Catalog and run with the caller’s permissions. Give each one an owner and a review step, and test it with a low-privilege user as well as an admin.
5. Keep benchmarking. Build a benchmark set for each Genie Agent and re-run it whenever the ontology, instructions or tables change. Learned context can drift, and a benchmark is how you find out.
6. Turn on tracing early. Unity Gateway can trace and audit what Genie did and why. You want that history from day one, not from the first time someone asks what happened.
7. Be careful with sharing. OpenSharing and Genie Clean Rooms make it easy to give a partner an agent. Review conversation and output limits, and what a recipient could infer from the answers, before you do.

The short version

• Genie can now run jobs on schedules and triggers and take actions with approval, not just answer questions.
• Genie Ontology, Genie Search and Genie MCP are GA according to Databricks’ companion posts. Ontology Manager is still coming. Release stages for autonomous jobs, Artifacts and Skills are not stated, so check the docs.
• The 91% benchmark came from a mature deployment. Your starting accuracy depends on your definitions.
• Write governed definitions, set Unity Gateway approvals and budgets, and start with read-only jobs before letting Genie act.

If you are planning a Genie rollout and want help with a readiness review, governance setup or a pilot, see our Databricks Genie implementation page.