Databricks's new Genie Ontology context engine is in gated preview, but it doesn't solve the core gap Thoughtworks flagged this week: agents stall not because platforms can't reach data, but because they don't understand what data means to the business. One client's catalog had 47 tables with "revenue" in the name. The agent could query all of them. It couldn't pick the right one, apply regional definitions correctly, or explain its choice.
Thoughtworks splits the failure into two layers. The platform layer—Unity Catalog, access controls, lineage—is mostly solved. The meaning layer is not. "Revenue" in a single company means booked, billed, or recognized depending on finance team. "Customer" means different things to sales and operations. Two tables sharing a customer_id column share a key, not necessarily a join. Thoughtworks recommends labeling relationships as inferred until a domain owner verifies them. That discipline prevents silent errors that surface only when an agent's output is checked after a decision.
Databricks offers three tools for the meaning layer. Genie Ontology, the bottom-up path, scans notebooks, dashboards, pipelines, and lineage, extracts knowledge snippets, scores them via OntoRank (a PageRank-style engine), and injects relevant ones at query time. It maintains itself—key because enterprise metric definitions drift after teams hand-build them. OntoBricks (Databricks Labs, no SLA) takes the top-down path: import standards like FIBO, FHIR, or CDISC, formalize as ontologies, materialize on the lakehouse, and run inference. Thoughtworks built OntoBricks into its regulated-domain accelerators. Ontos (also Labs) is a business catalog wrapping Unity Catalog with data products, contracts, and compliance rules—for organizations whose gap is ownership agreements, not technology. The three tools compose; they don't compete.
| Tool | Approach | Source / Standards | Maintenance | Primary Use Case | Availability |
|---|---|---|---|---|---|
| Genie Ontology | Bottom-up — scans artifacts | Notebooks, dashboards, pipelines, lineage | Self-updating (OntoRank scoring) | General-purpose semantic context at query time | Gated preview (account team approval required) |
| OntoBricks | Top-down — import formal standards | FIBO, FHIR, CDISC | Manual / standards-driven | Regulated domains (finance, healthcare, life sciences) | Databricks Labs — no SLA |
| Ontos | Governance wrapper | Unity Catalog data products & contracts | Ownership-agreement driven | Organizations whose gap is data ownership, not technology | Databricks Labs — no SLA |
Benchmarks from Databricks's internal 28-question suite (June 2026) showed Genie answering 84.5% correctly on the first attempt. The strongest general-purpose coding agent scored 52.4%; the weakest, 25%. Genie also ran 2× faster. Databricks anonymized competitors, so independent replication isn't possible yet, but results align with Thoughtworks production work: a governed context layer beats prompting a capable model against a schema it doesn't understand.
Success requires five organizational disciplines Thoughtworks identifies. Formalize metric definitions (formula, owner, dimensions, system of record, temporal rules, exceptions) in a shared platform layer—not prompts or agent code. Define business language and be honest about relationship status. Name what is trusted: the most-used dashboard isn't always correct. Attach evidence to every claim for defensibility months later. Version definitions alongside the learned layer, pin agents to a named release, and test against that release rather than live data. This addresses a risk tooling alone can't prevent: a self-updating context is right for discovery, not for CFO board presentations.
| Discipline | What It Means in Practice |
|---|---|
| Formalize metric definitions | Capture formula, owner, dimensions, system of record, temporal rules, and exceptions in a shared platform layer — not in prompts or agent code |
| Define business language honestly | Document what terms like 'Revenue' and 'Customer' mean per team; flag relationships as inferred until a domain owner verifies them |
| Name what is trusted | Explicitly declare the source-of-record table; the most-used dashboard is not always correct |
| Attach evidence to every claim | Record the basis for each definition so decisions remain defensible months later |
| Version definitions and pin agents | Release definitions alongside the learned layer; test agents against a named release, not live data |
Genie Ontology requires approval from the Databricks account team; it's not self-serve. OntoBricks and Ontos carry no SLA; treat them as proof-of-concept accelerators, not production commitments. The readiness gap is real and specific: if your organization can't answer "which table is the source of record for EBITDA and who owns it," the platform can't answer it either.