In this briefing
- Databricks has made Genie Ontology generally available. The product learns business knowledge from dashboards, SQL, notebooks and semantic models, while business teams manage the metrics, domains and concepts that require exact definitions.
- The second generation of Genie announced by Databricks can delegate long-running work to multiple subagents. DeepSeek Harness v0.2.2-rc.2 adds multi-level SubAgent collaboration, a shared task panel and Git Worktree review. The former is coming to preview and beta programmes, while the latter remains a pre-release.
- Databricks' vendor benchmark indicates that shared business context may reduce the cost of repeatedly finding rules. The IBM and Gartner material also requires enterprises to retain evidence of responsibility, trace, intervention and exit. Enterprises still need independent acceptance tests using their own data, permissions and failure conditions.
A shared ontology begins to carry changing business context
On 11 November, Databricks announced that Genie Ontology was generally available. The product learns business context from the dashboards, SQL queries and notebooks that teams use every day, as well as from semantic models in tools such as Power BI. The product stores calculations, business rules and data relationships as continuously updated Ontology Snippets. Business teams manage the metrics, business domains and authoritative concepts that require exact definitions through Metric Views, Domains and Pages, so automatically learnt knowledge and definitions confirmed by people can coexist.
Databricks says that OntoRank selects context using source authority, certification status, the expertise and relevance of the people involved, popularity and recency, while Genie Ontology continues to respect existing user permissions. Genie MCP also allows Agents outside Databricks to use the same context. The Databricks press release confirms that Genie and Genie Ontology are generally available, but the release lists Genie Search, Ontology Manager and the second-generation always-on Genie as coming to preview and beta programmes. The Genie Ontology blog describes Genie Search as generally available. Because the two official sources conflict on product status, this article does not treat the availability of Genie Search as a conclusion.
For enterprise adoption, a shared ontology cannot store only retrieval results. An enterprise needs to record the source, owner, version, validity period, scope, permissions and conflict-resolution rule for every business definition, and every Agent call needs to be traceable to the definitions it actually used. Automatic learning can broaden coverage, but it can also introduce an outdated dashboard, a copied notebook or an informal query into the context. Human certification, deprecation and rollback therefore remain necessary controls.
Long-running work connects shared meaning with isolated execution
Databricks describes the second generation of Genie as an always-on Agent that can run in the background, manage long-running work and delegate work to multiple subagents. Databricks says that Unity Catalog permissions and the approval rules, budgets and audit records in Unity Gateway constrain every action. The Databricks press release also states that the new always-on Genie is not generally available and is coming to preview and beta programmes, so the published feature description is not a production-service commitment.
On 11 November, DeepSeek released DeepSeek Harness v0.2.2-rc.2. This pre-release allows Git Worktree to review changes by branch, commit and file type, and it adds named and multi-level SubAgent collaboration, an AgentTeam that can be enabled for each session, and a shared member and task panel. The release also fixes repeated Agent wake-ups, task stopping and message retries. These fixes show that task state and stop semantics for parallel Agents remain engineering objects that are still being refined.
For enterprise adoption, multiple Agents can share business definitions, but they should not share execution state that cannot be distinguished. An enterprise needs to bind every task to an input version, working directory, permissions, budget, approval basis, stop state and output commit, and the enterprise needs to preserve idempotency when a subtask fails, an Agent wakes repeatedly or a network call is retried. The Databricks preview direction and the DeepSeek Harness pre-release provide inspectable design signals, but neither replaces production acceptance tests for permission leakage, duplicate tasks, rollback and human takeover.
A vendor benchmark compares business context at task level
Databricks and Turing created a fictional UK energy supplier, Calderwick Energy, for the benchmark. The test workspace contains hundreds of tables, including a largest table with around 120 million rows of half-hourly meter data, as well as notebooks, scheduled jobs, dashboards, saved SQL, code, metrics, documents and Genie Agent instances. The test also includes copied notebooks, abandoned analyses, old documents and development copies. The three Agents received the same questions, every task ran five times, and the grading criteria were defined before the runs began.
Databricks reports that Genie passed 91% of task runs, while Agent A and Agent B passed 45% and 58% respectively. Their average time per run was 100.3 seconds, 434.6 seconds and 306.5 seconds respectively, and their average token use was 431,829, 889,019 and 1,051,070 respectively. Genie ran natively, while the other two Agents connected through a Databricks toolbelt that did not include Genie MCP. The result comes from a vendor benchmark developed by Databricks and Turing, and it reflects an established deployment that has accumulated business knowledge. The result is not an independent test and does not show that every enterprise dataset will produce the same difference.
For enterprise adoption, business context can become infrastructure that requires its own acceptance test instead of remaining an unexplained prompt attachment. An enterprise can compare complete-task pass rate, total time, token use, tool calls, permission errors and human takeover on its own work, and it needs to test separately an established environment with accumulated knowledge, a cold-start environment and an environment in which definitions change quickly. The vendor benchmark supports the direction that shared context may reduce repeated discovery of rules, but an enterprise still needs to test whether incorrect context allows an Agent to produce an incorrect result more quickly.
Business meaning and Agent actions both need verifiable evidence
On 11 November, IBM published the Digital Sovereignty Index. The IBM research surveyed 2,003 senior decision-makers at large enterprises in the United States, United Kingdom, Germany, France, Saudi Arabia, India and Japan. IBM also conducted 25 policy-related interviews and collected views from 17 customer leaders. IBM reports that 94% of surveyed enterprises had faced at least one sovereignty challenge in the previous three years, 81% of the affected respondents had needed to provide formal evidence of control, and 44% of respondents identified artificial-intelligence-related issues as the most common individual challenge. IBM also reports that 82% of respondents had high confidence in providing evidence, but only 26% of surveyed enterprises reached the Index's leading tier and only 33% of respondents tested the relevant capabilities regularly or continuously.
IBM uses six dimensions — responsibility, trace, constraint, intervention, legitimacy and exit — to assess the control evidence reported by surveyed enterprises. On the same day, Gartner forecast that more than 90% of organisations in the European Union would have a digital-sovereignty strategy by 2029. Gartner also argues that a completely independent technology stack is not commercially attainable, so an organisation needs to identify unacceptable dependencies, retain alternative paths and manage interdependence. The IBM figures come from IBM research and respondent reports, while the Gartner percentage is an analyst forecast. Neither set of material replaces an audit of a specific enterprise.
For enterprise adoption, definitions in an ontology, the context used by an Agent, executed tools, human approvals, failure intervention and a supplier-exit plan need to form one evidence chain. An enterprise needs to prove who approved a definition, which Agent used that definition under which permission, how the system stopped or recovered a task, and whether the enterprise can export semantics and operating records when a supplier changes. An evidence chain cannot remove external dependencies, but it can make the scope of dependency, recovery capability and responsibility boundary inspectable.
What to watch next
- Can Databricks reconcile the public availability status of Genie Search, and can Databricks provide more complete version, conflict, deprecation and rollback records for ontology knowledge?
- Can DeepSeek Harness stabilise task idempotency, stop semantics, worktree isolation and cross-session auditing for multi-level Agents in a final release?
- Can enterprises reproduce the task benefit of business context with their own data and permissions, while also proving the handling of incorrect context, supplier interruption and exit recovery?
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing