Databricks unveiled a new version of its Genie agent on Wednesday, in a shift in who does the asking. Genie no longer waits for a prompt. According to TechCrunch, the updated agent takes ownership of multi-step jobs, delegates work to sub-agents, follows up on missing inputs, and returns only when a human decision is required.
The demo, fittingly, wasn't a demo. CEO Ali Ghodsi opened his own Genie and walked TechCrunch through his actual chats: the agent had analyzed the parts of Databricks' employee satisfaction survey where the company scored poorly, then rated — and helped discard — fixes executives had proposed. It summarized a sales pitch that went badly and prepped him for a podcast interview by reviewing past appearances.
The interesting part isn't the feature list. It's why a CEO would want an agent that tells him things his own company won't. Ghodsi invoked Dunbar's number — the idea that humans can maintain stable relationships with about 150 people. Databricks has 12,000 employees. At that size, nobody confesses a disastrous sales call or a doomed HR fix to the CEO. Genie, which can read the survey data, the tickets, and the call notes, has no such hesitation.
The autonomy rests on Genie Ontology, a context layer that went generally available with the launch, capturing a company's definitions and metrics so the agent knows what to trust. Genie Search, also generally available, indexes documents and tickets across Google Drive, Jira, Confluence, and Workday. It all runs inside the company's Databricks environment, governed by Unity Catalog and Unity Gateway, which lets admins trace Genie's actions and swap in cheaper open-weight models — Ghodsi's answer to what an always-on agent does to the AI bill.
The timing is no accident. Anthropic launched Claude's office suite — Sheets, Whiteboards, and a meeting participant — the same day. The AI coworker race is now two-front: Anthropic builds the apps, Databricks bets the winner owns the context underneath them. General Motors has already deployed more than 1,000 Genie agents, and Databricks' own benchmark, built with Turing, had Genie pass 91% of task runs vs. 58% for the stronger of two anonymized frontier agents — self-reported, but the direction is clear.
Genie One is free through January 31, 2027. The open question: whether workers trust an agent reading their surveys and Slack threads, and whether executives actually want the unvarnished truth on schedule.
— via TechCrunch