Fresh context for agents. Rebuilt every 12 hours. Provenance by Synorb.
Updated 2026-11-20 12:00 UTCnext update 2026-11-21 00:00 UTC
Reported · 4 Manifests890 words

Databricks bets data agents on governance: accuracy jumps when definitions are shared

Genie One accuracy jumps when curated agents sit on governed definitions, and Genie ZeroOps fixes pipelines only with a human sign-off. The pattern: control first, autonomy second.

Key facts
  • Databricks reports that curated agents and a governed layer each reached 83% accuracy independently, while the combined setup reached 93-95% accuracy. Self-reported m:…4093
  • Genie ZeroOps, now in public beta, triggers automatically when it detects a pipeline problem and creates pull requests for fixes, but Databricks says it will never change code or data without direct approval. Self-reported m:…4092
  • Databricks says ZeroOps was built from lessons in over 80,000 customer issues, and that over half of pipeline issues require organization-specific knowledge to resolve. Self-reported m:…4092
  • Genie Migrate for Serverless, in beta, automates moving classic workloads to serverless compute, proposing changes as reviewable diffs with deterministic validation and a reversible process. Self-reported m:…3947

The accuracy finding

Databricks has published research on its Genie One data agent that points to an uncomfortable conclusion for anyone betting on smarter agents alone. In the company's tests, curated agents and a governed layer each reached 83% accuracy independently, and the combination of the two reached 93-95%. The two parts solve different problems: adding curated agents improved document-based answers, while adding the governed layer improved answers that depend on business definitions. The reason this matters is that different departments often give contradictory but correct answers because they define the same term differently, and an agent that does not know which definition applies cannot tell which answer is right. Finance and Underwriting, the company says, are a case in point. m:…4093

The governed layer in question lives in Unity Catalog, where Databricks says governed assets like Metric View and Pages supply the business context agents otherwise miss. Genie Ontology captures business logic from existing dashboards, queries, and notebooks as Ontology Snippets, and updating a definition in a Metric View or Page propagates the change without editing individual agent instructions. That propagation is the practical payoff: one definition changes once, everywhere, instead of being re-taught agent by agent. The company also notes a gap worth knowing about. Genie One does not automatically inherit instructions configured inside an agent when it answers a question directly, which is why the company argues a unified governed layer is essential for consistency. In a retail test the governed layer raised accuracy from 54% to 82% when added to curated agents, and the insurance test used 8,000 policies and 7,482 claims. The tables-only configuration for Genie One achieved 64% accuracy, which is the baseline this whole argument is measured against. m:…4093

The agent that fixes pipelines

The second release is Genie ZeroOps, now in public beta, an autonomous agent that handles data pipeline failures from detection through to a proposed fix. Databricks introduced it in June, and the company says it monitors jobs, pipelines, and tables for failures, catching both hard job failures and silent data quality issues. When something breaks, it gathers context from logs, source code, table metadata, and lineage, then tests proposed fixes in a sandbox environment and sends alerts to Slack, Teams, or PagerDuty. The design is deliberately limited in one important way: the company says ZeroOps will never change code or data without direct approval, and it triggers automatically without a prompt, so the autonomy is in detection and diagnosis, not in rewriting production. m:…4092

The evidence base behind ZeroOps is large, and the company is candid about what it learned from it. Databricks says the agent was developed from insights in over 80,000 customer issues observed during private preview, and that over half of pipeline issues require organization-specific knowledge to resolve. That is the finding that explains why a general coding agent is not enough here: half the job is knowing how this particular organisation's pipelines are supposed to behave, which is exactly the kind of context that lives in lineage and metadata rather than in a model. This suggests the real product is not the fix itself but the investigation that precedes it, packaged in a form a human can approve. m:…4092

Migration, with the same guardrails

Genie Migrate for Serverless, in beta, applies the same philosophy to a different chore: moving classic Databricks workloads to serverless compute. The agent assesses workload readiness by scanning jobs and notebooks, translates infrastructure and code configurations, and proposes changes as reviewable diffs, with deterministic validation before anything is applied and a reversible process that keeps the original workload intact. There are limits the company states plainly: the agent cannot currently migrate workloads managed by Databricks Asset Bundles or Git folders. m:…3947

The framework that ties these releases together is Databricks' four-pattern guide for AI in regulated industries, which argues that enterprise AI cannot replace every deterministic control. The patterns are deterministic control with agentic reasoning, for outcomes that must be strictly reproducible; agentic execution with a deterministic judge, for tasks where results can be objectively tested; agentic reasoning with human authority, for judgment that cannot be proven by deterministic tests; and agentic execution with bounded autonomy, for agents acting within an explicit operating envelope. In the first pattern, agents can investigate control failures but cannot become the control, and in the third, human review must sit at the decision boundary. Unity Catalog governs enterprise data and enforces access rules including residency and cross-border transfer, while MLflow preserves traces and evaluation evidence for audit records. m:…3944

What this means for data leaders

The through-line is that Databricks is selling governance as the thing that makes agents safe to scale, and the numbers it reports support that framing at least on its own tests: accuracy rises from 83% to 93-95% when the governed layer joins the curated agent, and the fixes an agent proposes stay behind a human approval step. The honest caveat is that these are the company's own tests and its own products, so the accuracy figures should be treated as a claim to verify on your own data, not as a guarantee. The decision for a data leader is therefore narrower than the marketing suggests. Before scaling data agents, check whether your business definitions are written down in one place that agents can actually read, because that is the layer Databricks says does the work. If they are not, the agent will be guessing at definitions, and no amount of model quality will fix that. m:…4093 m:…4092 m:…3944

What is still unresolved

  • Whether the 93-95% combined accuracy holds on a customer's own data and definitions, outside Databricks' tests
  • How often ZeroOps' proposed fixes are accepted unchanged by the humans approving them
  • Whether the four regulated-industry patterns hold up under a real audit, rather than as a design guide

Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.

Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.

Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

  1. Get a Synorb Key. Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool synorb-get-demo-key (no header needed), or REST POST https://api.synorb.com/api/connect/demo-key with a JSON body such as {"client":"your-agent-name"}. The key is the api_key field of the answer. Save the key. Existing 72 Hour Agent Demo keys keep their issued limits. New keys require email OTP verification. An existing demo key lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons.
  2. Read each cited Manifest by id. REST: GET https://api.synorb.com/manifests/by-id/{manifest_id} with the header Authorization: Bearer <your Synorb Key>. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
  3. Check the claim. Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.
Cited Manifest ids (4)
  • 1795138047610134093
  • 1795138047610134092
  • 1795118971205873947
  • 1795118971204873944
for id in 1795138047610134093 1795138047610134092 1795118971205873947 1795118971204873944; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done

Provenance ledger

ManifestTitleSourceClaims usedPublished
1795138047610134093Genie One Accuracy: Why Curated Agents Need Governed Contextdatabricks-blog1795138221311135167, 1795138221312135168, 1795138221313135169, 1795138221306135164, 1795138221310135166, 1795138221315135171, 1795138221319135176, 1795138221316135172, 179513822131413517011-19 00:00
1795138047610134092Genie ZeroOps, now in public beta: An autonomous agent that fixes data pipelines, with lessons from 80,000 customer issuesdatabricks-blog1795138222299135255, 1795138222367135264, 1795138222368135265, 1795138222348135260, 1795138222361135261, 1795138222185135247, 1795138222192135248, 1795138222210135249, 1795138222217135250, 1795138222222135251, 179513822224313525311-19 00:00
1795118971205873947Introducing Genie Migrate for Serverlessdatabricks-blog1795119152316874934, 1795119152333874937, 1795119152334874940, 1795119152335874943, 1795119152324874935, 179511915233187493611-19 00:00
1795118971204873944Four AI Architecture Patterns for Regulated Industriesdatabricks-blog1795119152193874845, 1795119152196874847, 1795119152199874850, 1795119152206874852, 1795119152212874853, 1795119152224874855, 179511915222987485611-19 00:00

Related stories