Databricks Always-on Genie: an agent that works while you don't

Databricks just turned Genie from a chat box into a coworker that does not wait to be asked. On November 11 the company announced the always-on Genie, a second-generation autonomous agent. You give it a goal. It runs in the background on its own compute, hands pieces to sub-agents, chases missing inputs and pings a person only when a decision is needed. The pitch is clear. The fine print matters more: the always-on mode is not shipping yet, and the controls you set before it ships decide how it behaves.

What actually happened

From the Databricks press release:

• Autonomous mode. Genie "takes ownership of the work, delegates to subagents when needed, tracks progress, follows up on missing inputs, and reaches out when a person needs to make a decision."
• Priority inbox and email. Genie now opens to an inbox of items that need attention, and each new alert arrives as its own unread email.
• Genie everywhere. Through Genie MCP, teams can use it from ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot. New MCP connectors include Atlassian, Google Drive and GitHub.
• Governance. Every action stays inside Unity Catalog permissions and Unity Gateway approval rules and budgets, and is auditable. Admins pick which tools Genie can use, set approval rules and spending caps, and trace every action.
• Pricing. No list price. Databricks says customers "pay for what they use, not per seat," with spending limits by team or user.
• Availability. Genie Ontology is generally available. The always-on Genie, Genie Search and Ontology Manager are "coming soon to preview and beta programs."

Why it matters for your business

Most small teams will not run Databricks. The pattern still lands on your desk, because every major vendor is moving the same way: agents that work while nobody watches. Perplexity Teammates did it in Slack. Now it is the data platform.

An agent that waits for a prompt fails in front of you. An agent that runs overnight fails in a log. The difference between useful and expensive is three settings: what it can touch, what needs a human yes, and how much it can spend. Databricks puts all three in Unity Gateway. Good. But a default is not a policy, and usage pricing with no seat cap means a loop that runs all night bills all night.

What we would do before turning on any background agent, from any vendor:

1. Write the approval list first. Anything that sends, pays, deletes or changes a customer record needs a human yes.
2. Set a spend cap per user and per team on day one, below what would hurt.
3. Pick one recurring job, like a weekly revenue brief, and run it for a month before adding a second.

Sources: Databricks press release.