Databricks Turns Genie Into an Always-On Agent for Multi-Step Work Its CEO Already Uses Daily

Databricks on Wednesday unveiled an updated version of Genie, its AI agent for business data, upgraded from answering questions to completing multi-step jobs that take extended, unsupervised work — the company says — to finish. The connecting tissue is an ontology Databricks built to give Genie context from the information customers already store on its platform.

Rather than a canned demo, CEO Ali Ghodsi showed TechCrunch his actual production agent: he used it to analyze the sections of an employee satisfaction survey where the company scored badly, to rate and discard executives' proposed fixes, to summarize a sales pitch that went poorly including the features the prospect demanded before buying, and to prepare for a podcast interview by reviewing his past appearances and drafting likely questions. While the interview was underway, he asked Genie what he should say to the reporter — and conceded the advice was not bad.

The interesting part is what Ghodsi says he uses Genie for: hearing the things a CEO of 12,000 employees no longer hears directly. He invoked Dunbar's number — the anthropological limit of roughly 150 stable human relationships — to argue that past that size, a salesperson will not confess a disastrous call and an HR manager will not admit a failing program, and an agent with access to the real data can surface them anyway. Anyone with more than 150 employees can get value from this, he said.

As pitched, Genie follows what coding agents and harnesses have done for programmers, but for the business side of the company. Ghodsi said he also uses it in his personal life rather than turn to alternatives he considers security-questionable, naming OpenClaw, Instinct and Muse. Note that the claim of fully unsupervised completion is the company's own; independent benchmarks of Genie's multi-step reliability have not been published.

On the cost question — always-on agents are a way to burn API budget at scale — Ghodsi pointed to Unity Gateway, Databricks' usage-control product, which lets companies set spending limits and swap in cheaper open-weight models when the task does not need a frontier model. That is also a competitive answer to the concern that enterprise agents quietly ratchet up inference costs month over month.

The update lands amid a fierce enterprise-agent market: OpenAI, Microsoft, Salesforce and Anthropic are all shipping agent platforms aimed at the same workflows. Databricks' differentiation is that Genie reads the governed data lakehouse the customer already trusts, rather than asking companies to export their data to a chatbot — a structural advantage in enterprises where data residency rules decide vendor shortlists.