Databricks' Genie Beats Frontier Agents in 91% Benchmark Run

Databricks' Genie passed 91% of task runs in a benchmark built with Turing, versus 58% for the stronger of two anonymized frontier agents — and on November 11, 2026, the company used that number to relaunch Genie as an autonomous AI coworker for long-running enterprise work.

Genie can now run scheduled or event-triggered jobs, follow up on missing information and manage several assignments at once. Users decide which actions require approval. Databricks says thousands of customers already run Genie in production.

The new version takes responsibility for the work around an analysis: monitoring changes, requesting input and acting as new information arrives. Databricks' own examples include routing insurance claims that move outside underwriting guidelines and proposing marketing budget changes for approval.

What makes the autonomy possible?

The center of the launch is Genie Ontology, now generally available. It automatically builds reusable knowledge snippets from dashboards, SQL queries and notebooks. Unity Catalog Pages and Metric Views preserve explicitly governed definitions.

Databricks' OntoRank system weighs source authority, certification, user relevance, popularity and recency to select context and resolve conflicting definitions. Organizational directories, also generally available and starting with Workday, add reporting relationships and collaboration signals to help rank relevant knowledge.

In short: the ontology decides what Genie knows, and OntoRank decides what it trusts.

How far does the search reach?

Genie Search is generally available for Google Drive, Jira, Confluence and Workday, with SharePoint in beta. It indexes content from those systems and enforces their permissions before returning results. The index can combine with SQL and MCP connectors, which retrieve live information and take actions in external tools.

Not everything is GA. External BI knowledge sources are in beta, initially supporting Power BI. Ontology Manager, the tool for reviewing and curating the ontology, is coming soon in beta.

Who is actually running this?

General Motors has deployed more than 1,000 Genie agents across sales, engineering, supply chain and HR operations, Databricks reports. That figure illustrates the multi-department model the company is pursuing.

Genie Agents can now analyze PDFs, slide decks and other documents alongside governed tables in Public Preview, returning page-level citations. Native MCP connectors are generally available, letting agents create tickets, update records and post messages in external tools.

Two beta features extend the agent outside the company. OpenSharing provides a governed agent preloaded with data, instructions and knowledge across Databricks accounts. Genie Clean Rooms lets recipients ask questions without seeing the underlying data or proprietary logic, with conversation and output-size controls.

What keeps an autonomous agent in check?

Unity Gateway supplies the administrative layer. Traces capture the question, answer, cited objects, response time and actions — a queryable record of everything Genie does. Administrators can restrict MCP tools, require approvals, set budgets and spending caps, and review usage. Genie acts under each user's identity and can prepare an access request, but administrators and data owners keep approval authority.

Genie One, the interface for this work, also builds slide decks, dashboards and interactive pages, with scheduled refreshes and shared skills for repeatable workflows. It's free for users through January 31, 2027, with more detailed cost controls promised before that date.

What does the benchmark actually prove?

Each agent in the Turing-developed evaluation ran every task five times in the same workspace: a fictional UK energy supplier containing tables, notebooks, dashboards, saved queries and documents — including outdated and duplicated assets. Genie averaged 431,829 tokens and 100.3 seconds per run.

Databricks itself cautions that the test reflects an established deployment with accumulated business knowledge. Differences may shrink in new deployments or when table schemas alone suffice.

That caveat is the real story. Broader autonomy depends on governed, accessible institutional knowledge — and on administrators deciding which jobs and actions Genie can take on.