Databricks announced Genie ZeroOps, an autonomous background agent designed to monitor production data and AI workloads, investigate issues, and suggest fixes that teams can verify before applying.
The new agent is built into the Databricks Platform and is intended to help data teams reduce the time spent maintaining production pipelines, jobs, tables, machine learning models, and related assets.
Databricks said data and AI teams are spending too much time responding to operational issues, including broken pipelines, upstream schema changes, late-arriving data, silent data quality problems, and machine learning model drift. The company said the rise of large language models and agentic development tools has made it faster to ship pipelines and models, increasing the need for automated operations support.
Genie ZeroOps works by continuously monitoring data and AI assets, detecting failures or data quality issues, assessing root causes using Unity Catalog lineage, generating proposed fixes, and validating them in a secure sandbox before production changes are made.
Because the agent runs within Databricks, it can access platform observability data, including metrics, events, logs, run history, and lineage information, while operating under Unity Catalog governance. Databricks said this allows Genie ZeroOps to trace problems to their underlying source, including code bugs, upstream schema changes, or bad data introduced by another pipeline.
The platform uses sandbox environments with zero-copy shallow clones of production data, scoped permissions, and network isolation. This allows proposed fixes to be tested against real data without touching production systems.
Databricks said Genie ZeroOps is designed for data and AI operations rather than general coding assistance. The company noted that coding agents can help write software, but typically lack access to telemetry, lineage, governed production data, and safe validation environments needed to detect, diagnose, and verify fixes for data and AI workloads.
For machine learning workloads, Genie ZeroOps can diagnose issues when models continue running but produce degraded predictions. The agent can build corrected candidates, evaluate them against the same evaluation suite used for the production model, and surface replacements only when they perform measurably better.
Users will be able to configure which assets Genie ZeroOps monitors and what actions it is authorized to take. Issues appear in an inbox-style interface prioritized by severity, with root cause analysis and proposed fixes. Databricks said nothing is applied to production without user approval.
Genie ZeroOps is entering private preview in the coming weeks, starting with support for jobs, pipelines, tables, and machine learning workloads. Apps and Lakebase databases are on the roadmap.
