Databricks Upgrades Genie, Its AI Agent, for Multi-Step Work
Databricks has released an updated version of Genie, its AI agent, that can carry out multi-step jobs requiring extended work. TechCrunch reported the launch on November 11, 2026, and the company is positioning the agent as an always-on tool for business leaders.
Genie draws its context from an ontology that Databricks built: a structured map of entities, such as people, products and customers, and the relationships between them. The ontology lets the agent reason over information that customers store on the Databricks platform, rather than treating that data as an undifferentiated pile of tables and files. Think of it as an org chart that also records how every team, product and record connects to the others. The company's pitch rests on that grounding layer.
Databricks CEO Ali Ghodsi uses Genie internally, and in place of a conventional product demo he shows his actual agent chats and charts. Among the tasks he has run through it, he has used Genie to analyze the lowest-scoring sections of an employee satisfaction survey. He has also used it to rate fixes that executives proposed, keeping some and discarding others. In a separate case, he asked the agent to summarize a sales pitch to a prospect that went badly, and the summary surfaced features the prospect said it needed before it would buy.
Ghodsi describes Genie as a source of unvarnished truth about what is not going well, the kind of information he is unlikely to receive directly as chief executive. Databricks has 12,000 employees, and he says that at that size he cannot know everyone. The agent, in his framing, fills a gap that organizational scale creates.
On cost, Databricks says agent AI spending can be controlled through its Unity Gateway product. According to the company, Unity Gateway allows administrators to set usage limits and to swap in less expensive open-weight models, meaning models whose trained parameters are publicly released, for particular workloads. The claim addresses a practical concern for teams running agents that make many model calls per task, where token consumption on long, multi-step jobs can grow quickly. The company has not, in the reporting, tied that claim to specific savings figures, so the cost argument rests on the control mechanisms themselves.
In my view, the more interesting part of this story is the choice of demonstration. A product demo shows what a tool can do under controlled conditions. Showing a CEO's own chats and charts shows how the tool handles questions the CEO did not script, including the uncomfortable ones. That is a credible way to make a case to technical buyers, who tend to discount polished walkthroughs.
The broader context here is that the agent's value depends on what it can see. An ontology-grounded agent is only as candid as the data it is permitted to read, and the survey and sales examples both hinge on information that employees and prospects supplied in good faith. Governance questions, such as who can query what and how agent outputs are audited, follow directly from that dependence.
Still, the underlying use case is a plain one. A leader with thousands of people and too many channels has long struggled to get an unfiltered read on the organization. If an agent can reliably sort survey comments, grade proposed remedies against evidence, and replay a lost deal in terms of what the buyer actually asked for, the executive gets a faster and less filtered view of the business. Whether that holds up beyond a single CEO's workflow is the question the next round of customer deployments will answer.
