Databricks has widened what its Genie Agents can work with: thousands of documents sitting beside database tables, connectors to workplace tools, and a way to hand a finished agent to a customer or partner. The company laid out the changes in a blog post by Hanlin Sun and Nikhil Gangaraju. Genie Agents launched as a tool that answered business questions by writing SQL. This release makes them a general business agent that reads, acts and reaches outside the company.
Release status differs by feature. Document analysis is in Public Preview. The connectors, built on MCP (Anthropic’s protocol for tool calling), have reached general availability, while sharing through OpenSharing and Genie Clean Rooms is in beta. Only the connectors are ready for a production commitment today.
The document capability is the largest change. Builders point an agent at a Unity Catalog volume holding PDFs, presentations, Word files or images. For each question the agent pulls the most relevant passages, reasons over them alongside table data, and cites its sources, sometimes down to the page. Databricks’s own parsing function reads tables, figures, headers and footers, so a single agent can search thousands of files for the ones that bear on a question. Because retrieval follows Unity Catalog permissions, users see only the files they can already open.
Databricks frames this against the usual alternative. Without a product like this, teams stitch together their own retrieval stack: a document parser, a search endpoint, an embedding pipeline, and a supervising agent layer that has to carry permissions along. The post says a homegrown version quickly becomes costly to build and to keep running. Genie Agents bundle those steps into one product.
The examples show the intended use. An investment analyst watches a portfolio company’s margin slide for three straight quarters. The agent pulls the product-line breakdown from governed tables, then searches analyst reports and earnings call transcripts for competitor moves. A technician gets repair steps out of thousands of service manuals, with tables and numbered steps kept intact. A clinical trial analyst spots a department whose study count is dropping, then has the agent read dozens of doctor reports to find out why.
On the action side, Lakeflow Connect ingests data from Salesforce, Workday, SAP and Marketo into Unity Catalog. The MCP connectors read from Slack, Google Drive, Google Calendar, Gmail, Glean, Atlassian, GitHub and Microsoft 365, and they can create tickets, update records and post messages. Agent authors choose which tools each agent may touch, and Unity Gateway is the control point for every connection. In Databricks’s sales example, an agent combines pipeline data with account plans and win-loss reports, then posts a summary to Slack, with the connector limited to posting summaries only.
The sharing features carry the most commercial weight. With OpenSharing, a provider can send a working Genie Agent, preloaded with instructions and context, to a recipient on a different Databricks account. Genie Clean Rooms go a step further. The recipient chats with the agent but never sees the raw dataset or the provider’s IP, and the provider can cap conversation length and output size. Databricks says data partners could list packaged agents in its Marketplace next to FactSet and LSEG. That points to a shift from selling datasets to selling answers, where the buyer never holds the data.
Evidence for the new document feature is thin. The only customer comment tied to it comes from Verana Health, whose engineering director says the team is looking to expand into unstructured clinical data, a statement of plans rather than results. The headline customer figures, HelloFresh’s 4x faster ad testing and General Motors’ more than 1,000 deployed agents, describe the existing structured-data product. The post offers no accuracy measurements for document answers. It instead describes benchmarks that customers write themselves, scored like unit tests against gold-standard SQL answers.
Teams already on Databricks should run the Public Preview against their messiest real document set and check the page-level citations by hand before connecting any of it to an agent that can post, file tickets or update records.
Announced by Databricks in a company blog post by Hanlin Sun and Nikhil Gangaraju.
