Databricks gave an always-on agent the keys to enterprise work on November 11, and gave it an agentic ontology to steer by. Genie Ontology reached general availability on November 9, according to Databricks release notes. The second generation of Genie, announced the next day, takes ownership of a goal, delegates to subagents, tracks progress and follows up on missing inputs, according to the Databricks announcement. The same announcement sets the limit. Genie and Genie Ontology are generally available. Genie Search, Ontology Manager and the new always-on Genie are coming soon to preview and beta programs. The map is generally available. The agent that acts on it, and the manager that lets people curate it, are not.

What the agentic ontology actually holds

Databricks defines Genie Ontology as a context layer that combines modeled knowledge a company defines in Unity Catalog semantics with learned knowledge Genie extracts from its assets and usage, per the release notes. The announcement adds the contents: definitions, metrics, people and processes, kept current as the business changes, learned from dashboards, queries, notebooks, external business intelligence tools and an organizational structure connector.

Chief executive Ali Ghodsi framed the test in the announcement: Genie works because it knows "what we mean by revenue, which numbers are official, and who can see what."

Genie Search expands that map to more than 100 enterprise data sources, including Google Drive, Jira and Confluence, and Genie runs inside a customer Databricks environment under Unity Catalog and Unity Gateway, with actions described as fully auditable and subject to approval rules and budgets, per the announcement.

Analysts quoted by TechTarget on November 11 made the same point from outside. Sanjeev Mohan of SanjMo said agents "fail less from bad models than from missing context." Mike Leone of Moor Insights and Strategy said Genie "stops waiting to be asked."

The numbers behind the context claim

Databricks released an evaluation developed with Turing, reported by Unite.AI on November 11. Genie passed 91 percent of task runs, against 58 percent for the stronger of two anonymized frontier agents, while averaging 431,829 tokens and 100.3 seconds per run. Each agent ran every task five times in one workspace for a fictional United Kingdom energy supplier, with tables, notebooks, dashboards, saved queries and documents, including outdated and duplicated assets. Databricks cautions that the test reflects an established deployment with accumulated business knowledge, and that the gap may be smaller in a new deployment or where table schemas alone are enough. The figure is a vendor evaluation, not an independent replication. Unite.AI also reports Databricks figures of thousands of customers using Genie in production and more than 1,000 Genie agents deployed at General Motors across sales, engineering, supply chain and human resources.

Independent survey evidence points the same way, with a caveat. MIT Technology Review Insights, in partnership with Snowflake, surveyed 400 senior data and technology executives in July and August 2026 and published on November 2. In that survey, 55 percent cited a lack of business context as the main reason agent insights fall short, 45 percent cited fragmented or inaccessible data, and on average only 48 percent of agent insights led to actions that improved the business. Insights is a sponsored research arm and Snowflake sells the semantic and governance products the findings support, so the survey is a market signal, not an audit of any one company record.

The agentic ontology test: who can overrule the map

The launch leaves one question open. Unite.AI reports that a system called OntoRank weighs source authority, certification, user relevance, popularity and recency to select context and resolve conflicting definitions. A ranking can pick a winner. It does not record who approved that winner, or what it replaced.

That distinction matters because the learned half of the ontology is inferred from usage. A definition that is popular and recent is not always the definition finance approved. Ontology Manager, the experience Databricks built so teams can manage and govern that context and refine frequently asked concepts, is the human review surface for the map. It is also the part still heading to beta, per the announcement and TechTarget, which reports it was developed in response to customer requests.

Mohan made the remaining gap explicit to TechTarget: enterprises need to govern what agents do, not only what they read, with a clear identity for each agent, limits on actions, human approval for consequential steps and a full audit trail. Leone added a second gap: a way for an always-on agent to say when it is not sure, because a wrong answer delivered unprompted can reach many people before anyone questions it.

The test for any agentic ontology is therefore a receipt test, in four parts. First, mark each entry as modeled or learned, with its source and date. Second, name the person or body that approved a definition when sources conflict, and keep the superseded version visible. Third, propagate a correction to derived summaries, standing instructions and queued always-on jobs, not only to the next answer. Fourth, require the agent to flag low confidence instead of delivering an uncertain result unprompted. That is the same standard set out in How Jilu Works and in the knowledge compilation and context retrieval posts: a retrieved or inferred fact becomes a record only when its source, date and superseded version can be inspected.

Databricks has shipped the strongest version yet of a claim the whole market now makes, that an agent is only as good as the company map under it. The map reaching general availability settles the existence question. It does not settle the authority question. Until the curation surface is as available as the ontology itself, and until a ranked definition carries the name of who approved it, an always-on agent is acting on a map that people can use but cannot yet fully overrule. For Jilu, that is the dividing line between a context layer and a company record: a record is the layer a person can open, dispute and correct, with the correction reaching every agent that read the earlier version.