Databricks unveils always-on Genie agent, coming soon to preview
Databricks announced the second generation of Genie on November 11 at Data + AI Live, announcing a second generation of Genie that will turn its data assistant from a question-answering chatbot into an always-on autonomous agent that runs jobs in the background, delegates to sub-agents and follows up on missing inputs — with the always-on mode coming soon to preview and beta programs. In a benchmark built with Turing, Databricks says Genie passed 91% of task runs versus 58% for the stronger of two anonymised frontier agents — its own evaluation, with its own caveat that the gap may be smaller in new deployments.
📊 Scorecard: our Bullish call on $MSFT is scored from $521.13, the last close before it ran. Its first result comes after the next close.
The News
Databricks announced the second generation of Genie on November 11 at Data + AI Live, that will turn its data assistant from a question-answering chatbot into an always-on autonomous agent for enterprise work. The old Genie answered questions about a company's data when asked. The new one, once the always-on mode reaches preview, works in the background: it runs scheduled or event-triggered jobs, delegates pieces of work to sub-agents — smaller agents that handle one task each — follows up when inputs are missing, and surfaces decisions to people through a priority inbox and email. Users decide which actions need human approval and which the agent can take on its own. Databricks is a private company, so the public read-throughs are Microsoft, whose Copilot Cowork reaches Genie users, and Snowflake, whose Cortex Agents are the rival lakehouse agent platform.
Genie and Genie Ontology — the context layer that captures a company's business definitions, metrics and processes, so the agent knows what 'revenue' means at that company — are generally available. Genie Search, with more than 100 connectors including Google Drive, Jira, Confluence and Workday, plus Ontology Manager and the always-on mode itself, are coming soon to preview and beta programs. Every action is governed by Unity Catalog and Unity Gateway, Databricks' control layers, and is auditable; Genie runs inside the customer's existing Databricks environment rather than sending data elsewhere. Genie One, the interface, is free for users through January 31, 2027.
Databricks says thousands of customers already use Genie in production and names General Motors, with more than 1,000 Genie agents deployed across sales, engineering, supply chain and HR. The company also released a benchmark built with Turing, a data-labelling and evaluation firm, in which Genie passed 91% of task runs versus 58% for the stronger of two anonymised frontier agents. The figure is Databricks' own evaluation, and the company itself cautions that the test reflects an established deployment with accumulated business knowledge — the gap may be smaller in a new deployment that has not yet learned the company's definitions.
The value of an enterprise agent is not in the answer — it is in the follow-through.
Why It Matters
The shift from answering questions to doing work is the whole story of enterprise AI in 2026. A chatbot that answers 'what were last quarter's sales?' saves a few minutes. An agent that notices a supply-chain delay, checks the numbers, drafts a response and asks a human to approve it saves a workflow. Databricks is betting that the value of an AI assistant is not in the answer but in the follow-through — and that companies will pay for an agent that keeps working when nobody is watching.
The 91%-versus-58% benchmark is striking, but it needs its caveat read alongside it. Genie was tested in an established deployment, where it had already absorbed the company's definitions, metrics and processes — its home turf, in effect. The two anonymised frontier agents were general-purpose models without that accumulated context. Databricks deserves credit for saying so plainly. The honest read is that a specialised agent with deep business context beats a general agent without it — which is a real insight, and a narrower claim than 'Genie beats the frontier.'
For the market, the read-throughs are Microsoft and Snowflake. Genie reaches users inside Microsoft's Copilot Cowork, so its success is also a win for Microsoft's agent ecosystem. Snowflake's Cortex Agents are the direct rival: two lakehouse platforms — systems that store and analyse a company's data — now competing on whose agent does the work better. Databricks is private, so the stock-market signal is indirect, but the direction is clear: the lakehouse agent war is now a product war, not a slide war.
Analyst's Take
Our call is Bullish. The always-on Genie Databricks announced — an agent that runs jobs, delegates to sub-agents and follows up on missing inputs — is a genuine step past the chatbot — and General Motors running more than 1,000 Genie agents is the kind of production evidence that matters more than any benchmark. The 91% figure comes with an honest caveat from Databricks itself, which makes it easier, not harder, to take seriously.
What to watch: whether the always-on mode and Genie Search move from preview to general availability, and whether other large customers publish their own Genie numbers. A thousand agents at one carmaker is a proof point; a dozen carmakers would be a trend. And watch Snowflake's answer — Cortex Agents is the rival, and this is now a race.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is Genie?
Genie is Databricks' AI assistant for working with company data. The first version answered questions when asked; the second generation, announced November 11, will work on its own in the background once the always-on mode reaches preview — running jobs, delegating to smaller sub-agents and following up when information is missing.
What is a sub-agent?
A sub-agent is a smaller AI agent given one piece of a bigger job — for example, checking inventory while the main agent drafts a report. Genie breaks work down and hands the pieces to sub-agents, then pulls the results together.
What is Genie Ontology?
It is the context layer that teaches Genie what a company's words mean — its definitions, metrics and processes — so that when someone asks about 'revenue,' the agent knows exactly which numbers count. It is generally available now.
What did the 91% benchmark actually test?
Databricks built a benchmark with Turing in which Genie passed 91% of task runs, versus 58% for the stronger of two anonymised frontier agents. But Databricks cautions the test used an established deployment with accumulated business knowledge — the gap may be smaller in a brand-new deployment.
Why are Microsoft and Snowflake the tickers?
Databricks is private, so there is no Databricks stock. Microsoft is the read-through because Genie reaches users inside Microsoft Copilot Cowork; Snowflake is the rival, because its Cortex Agents compete directly with Genie as the lakehouse agent platform.
👁
Source: Databricks