The report
An enterprise agent that continues working after its requester closes the chat needs more than good document retrieval: it has to know whether the original authorization still applies when the next action occurs. Databricks' November 11 announcement describes a second-generation Genie that takes longer-running objectives, delegates to subagents, monitors conditions and brings decisions back through a priority inbox and threaded email. Those features were announced as a coming preview, not a verified generally available autonomous service.
Three related components serve different functions. Genie Ontology represents business definitions, metrics, people and processes; Genie Search seeks governed information across lakehouse data and external business applications; the proposed always-on Genie would use that context to carry work forward across time. Databricks describes more than 100 connector options and three retrieval approaches: ingesting material, indexing content where it remains, or querying external systems through live MCP connections. The data location and update mechanism matter because a cached answer can remain available after its original access conditions change.
Each layer changes a different boundary
| Component | Value proposed by Databricks | Independent control question |
|---|---|---|
| Genie Ontology | Shared definitions and organizational context used to interpret work | Who can modify a decision-relevant definition, and when does that change reach running tasks? |
| Genie Search | Retrieval across tables, documents, tickets and connected apps | Are source permissions checked again when indexed or cached content is reused? |
| Unity Catalog and Unity Gateway | Data access, tool approvals, usage budgets and audit trails | Which identity executes a delegated operation, and how is its effective authority withdrawn? |
| Always-on execution | Background tasks, delegated subagents and asynchronous notification | What happens to queued work when the user, target system or approval state changes? |
Databricks asserts that Genie runs inside the customer's existing Databricks environment and follows its governance rules. The announcement does not provide a controlled revocation test showing that a child agent stops using an already fetched document, that an external tool cannot complete a dispatched change after a freeze, or that a new data permission propagates consistently across all three retrieval strategies. Those are specific falsifiable questions, not evidence that such failures have occurred in the product.
Availability and productivity need separate denominators
The vendor states that existing Genie and Genie Ontology are generally available. Genie Search, Ontology Manager and the new always-on mode were described as forthcoming preview or beta features. A quoted HelloFresh example reports 75 percent less time identifying effective Meta advertisements in one workflow; it is not a controlled estimate of savings across all customers, nor proof of policy compliance for autonomous tasks. The pricing description is usage-based, with Gateway spending limits, not a per-worker security guarantee.
An operator evaluating the eventual release should follow one real task from a named person's approval through retrieval, delegated tool calls, changed permissions, external effects and final notification. The DeputyBench permission study is not a Databricks test, but it demonstrates the general danger of treating a successful workflow description as an adequate account of actual authority. The distinguishing evidence would be a trace of the effect made under a still-valid grant, not merely a correct-looking final report.