Databricks Turns Genie Into an Autonomous Agent That Runs Enterprise Work

Databricks has introduced a new Genie release that expands the product from conversational analytics into a long-running enterprise agent. The company calls it an “autonomous coworker” that can monitor connected systems, respond to schedules or data changes, and act through approved tools without waiting for another prompt.

The product targets recurring, data-heavy work in which teams track metrics, investigate changes, and coordinate a response. Databricks combines the agent with business definitions, data relationships, permissions, and tools so it can interpret company data consistently and operate within policy.

Genie takes a standing assignment

Databricks illustrates the model with an insurance workflow in which Genie monitors incoming claims for emerging fraud patterns and loss-ratio drift, assembles a case file, and routes it to underwriters when a segment moves outside guidelines. In marketing, the agent can monitor campaigns, investigate underperformance, and propose weekly budget changes for approval.

A business map keeps terms stable

Genie Ontology, now generally available, supplies the structured business context behind the agent. It includes metric definitions, entity relationships, organizational context, and records of which numbers are authoritative. This semantic layer helps prevent an agent from assigning a new meaning to terms such as “active customer” during each task.

Databricks says agents bootstrap most of the ontology from dashboards, queries, and notebooks that employees already use, then incorporate signals from later queries and AI interactions. Teams can curate the resulting map by establishing authoritative definitions in Unity Catalog. An algorithm called OntoRank gives those definitions precedence when agents encounter competing sources.

In a benchmark conducted with Turing, Genie answered correctly in 91% of task runs, 1.6 times the success rate of a frontier-agent baseline, while using less than half as many tokens. The tasks required combining information across tables, dashboards, and notebooks. Because Databricks participated in the evaluation, the results establish performance for the published test design; teams still need tests covering their own schemas, permissions, and workflows.

Connectors widen Genie’s reach

• Genie Search: A hybrid retrieval layer combines indexed content from Google Drive, Jira, Confluence, Workday, and SharePoint with live Model Context Protocol lookups. The index handles fast retrieval, while MCP fetches recent updates and content outside the index.
• Genie Agents: Specialized agents can combine structured tables with PDFs and slides. They search content in Unity Catalog Volumes without requiring teams to provision a separate vector index.
• Genie plugins: Plugins let external assistants, including ChatGPT, Claude, Microsoft Copilot, and Grok, call Genie for governed business context.
• MCP connectors: Read and write integrations connect Genie to Slack, Zoom, Microsoft 365, Jira, GitHub, ServiceNow, PagerDuty, and calendar systems.
• OpenSharing and Clean Rooms: These features let partners use governed agents or query shared data while keeping the underlying tables and logic concealed.

Unity Gateway sets the boundaries

Unity Gateway provides the control plane for access, spending, and audit. Genie acts under the user’s identity and can access only resources available to that identity. Administrators can set workspace and per-user spending caps, while action traces provide a record for audits and incident reviews.

Governed context is Databricks’ wager

OpenAI, Anthropic, Microsoft, and Salesforce are also developing agents that perform long-running work across enterprise systems. Databricks places its advantage in the governed context layer between models, company data, and operational tools. Organizations already using Unity Catalog can reuse existing definitions, permissions, and lineage instead of rebuilding those controls in a separate agent framework.

That integration also couples agent behavior more closely to Databricks’ metadata and governance stack. Organizations using other data platforms can treat the design as an evaluation pattern: autonomous work requires stable business definitions, current source data, least-privilege access, human approval points, audit trails, and cost controls.

Checks before granting write access

• Availability and cost: Confirm which components are generally available or in preview for the required region and plan, then model compute, token, storage, and connector charges.
• Identity lifecycle: Determine what happens when a job owner changes roles, loses access, or leaves the organization.
• Write safety: Restrict each job to the systems and actions it needs, with approvals for irreversible, sensitive, or high-cost operations.
• Failure handling: Define timeouts, retries, idempotency, rollback behavior, and the response to duplicate or stale triggers.
• Semantic quality: Test conflicting metric definitions, incomplete lineage, stale dashboards, and malformed source data.
• Production evaluation: Measure task success, incorrect actions, latency, intervention rate, and total spend on representative internal workflows.
• Operational control: Verify trace retention, alerting, credential revocation, job suspension, and emergency shutdown procedures.