Databricks' Genie ZeroOps Hits Public Beta After 80,000 Issues
Databricks opened Genie ZeroOps to public beta on November 19, 2026, after the autonomous pipeline agent observed more than 80,000 real customer issues in private preview.

Why it matters
- Databricks opened Genie ZeroOps to public beta on November 19, 2026, after a private preview that observed more than 80,000 real customer issues.
- More than half of preview issues required organization-specific knowledge to triage, rank and resolve properly.
- ZeroOps requires explicit user approval before pushing any Git-backed fix; the push adds a commit without force-pushing.
- The agent runs Anthropic and OpenAI models on Databricks-hosted endpoints inside the company's security perimeter.
- The beta is workspace-level, not HIPAA compliant, and does not monitor models or apps.
Databricks opened Genie ZeroOps to public beta on November 19, 2026, after an autonomous data-pipeline agent spent a private preview observing more than 80,000 real customer issues across hundreds of companies. The company says that exercise produced a specific finding: more than half of those issues required organization-specific knowledge to triage, rank and resolve properly. That number shaped the beta's feature set, which now lets teams encode their own context into how the agent works.
ZeroOps detects failures, investigates their causes and prepares fixes for data infrastructure. The public beta adds shared inboxes, notifications that summarize each issue and recommendation, custom instructions for company-specific workflows, and one-click pull-request creation for Git-backed jobs. The stakes for enterprises are straightforward. Data teams lose hours to pipeline triage, and Databricks is betting that an agent can compress that diagnosis work while leaving production authority with the customer.
What did the private preview reveal?
The limited preview covered hundreds of companies before the November 19 opening. Databricks reports that the majority of observed issues — the company quantifies it as more than half — could not be properly handled with generic operational knowledge alone. They required organization-specific context.
That finding drove the beta's customization features. Teams can now identify known flaky sources, preferred fixes and naming conventions. They can filter expected failures and suppress lower-priority notifications. In effect, the agent's usefulness scales with how much institutional knowledge a team feeds it.
How does the agent actually work?
ZeroOps watches for two categories of problems. The first is explicit job and pipeline failures. The second is subtler: tables that stop updating, or tables that arrive with unexpectedly few rows — problems that may never trigger a conventional job failure alert.
Once it detects an issue, the investigation starts automatically. The agent gathers logs, source code, table metadata and lineage. It runs SQL queries to identify root causes. It then prioritizes its findings by downstream impact, so the team sees which failure is actually damaging the business rather than which failure happened to fire first.
Built-in access to run history, data-quality signals and Unity Catalog lineage — the Databricks service that maps dependencies between assets — connects code diagnostics to the state of the data itself. The connection matters. A code error means little without knowing which downstream tables consumed the broken output.
Proposed fixes do not go straight into production. ZeroOps tests them with unit tests and dry runs in an isolated sandbox before anyone applies them.
What are the approval guardrails?
Databricks has drawn a firm line on autonomy: ZeroOps does not change production code or data without approval. The documented guardrails require an explicit user instruction and confirmation before the agent pushes a Git-backed fix. The push adds a commit without force-pushing or rewriting branch history, so the repository's record stays intact.
There is a nuance that Git administrators should note. Naming a repository's default branch in the configuration can bypass pull-request review entirely, unless branch protection blocks direct pushes. Teams that rely on mandatory PR review need to check their branch settings before enabling the feature.
Alerts route to Slack, Microsoft Teams, PagerDuty or generic webhooks. Email destinations are unsupported.
What happens to sensitive data?
The agent runs on Anthropic and OpenAI models, hosted on Databricks-hosted endpoints inside the company's security perimeter. That architecture answers a common enterprise objection — the models do not sit outside the platform.
But the authorization model is broad. Selecting assets for ZeroOps authorizes continuous, autonomous large-language-model processing of those assets until the feature is disabled, including repository contents for Git-backed jobs. This is not a per-run approval. It is a standing grant.
Databricks has built in a masking step for external channels. Before a notification includes a finding's title or summary, the agent checks for sensitive information such as names and email addresses and masks what it finds. One caveat follows the data out the door: copies stored in outside systems fall outside Databricks' retention and deletion controls.
Who can see what?
A ZeroOps configuration can be shared with up to 25 individual workspace users. It cannot be shared with groups or service principals. Collaborators can work on the agent's findings together.
The visibility model has a sharp edge. Shared users see code excerpts, suggested fixes and sample data rows even without access to the underlying assets. Sharing a configuration does not grant access to those assets themselves, but the diagnostic material attached to them travels with the share. Administrators assigning permissions should treat a shared configuration as a partial window into assets a user cannot otherwise open.
What does setup require?
An administrator must enable Genie ZeroOps from the workspace Previews page. Users then select assets and need OWNER, CAN MANAGE or CAN VIEW permission on them. On startup, the setup process backfills findings from the previous 24 hours before reporting new issues, so the agent starts with recent history rather than a blank slate.
Git access splits by task. Investigating a Git repository requires a credential with read access. Creating a pull-request branch requires write access.
What is out of scope?
Databricks introduced ZeroOps on June 16, 2026 with planned private-preview support for jobs, pipelines, tables and machine-learning workloads. The current documentation narrows that ambition for the beta: monitoring covers jobs, pipelines and tables — not models or apps. Performance and cost regressions fall outside its detection scope.
Three further limits define the beta's enterprise posture. It operates at workspace rather than account level. It is not HIPAA compliant, which rules it out for covered healthcare workloads. And detecting late-arriving or incomplete table data requires separately billed data-quality monitoring — that capability is an add-on, not a bundled feature.
Why the trade-off matters
The beta's central enterprise bargain is less manual diagnosis in exchange for continuous AI processing of selected operational data. Databricks frames the resulting questions — sharing visibility, external notifications and Git review — as part of deployment itself rather than separate administrative details to sort out later.
That framing is the honest one. An agent that reads repository contents continuously, sends summaries to Slack and drafts fixes cannot be bolted onto existing governance as an afterthought. The 80,000-issue preview showed that generic automation covers less than half the problem; the beta's customization features are the response. Whether teams find that trade worthwhile will determine if ZeroOps moves from beta to standard practice for data operations on the platform.
Source: unite.ai
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
791 articles
Related articles
- Databricks Maps Three Enterprise Paths for OpenAI Tools on Governed Data
- Databricks turns Genie into an autonomous enterprise AI coworker
- Databricks launches always-on Genie AI agent for enterprise work
- Salesforce and Microsoft launch Copilotforce to bring CRM into Copilot
- Floqer passes $1.2M ARR and will open-source its AI prospecting database