
Image: Databricks
Why it mattersData teams get help on the slow part of a page at 3am. The agent will not push a change without a person approving it.
A pipeline fails at 3am, and the person on call spends an hour before they find the broken row. Databricks says Genie ZeroOps is now in public beta. The agent does the search work and drafts a fix as a pull request. A person still clicks merge.
Jeff Chou, Kristóf Molnár, Lennart Kats, Danny Chiao and Ray Zhu wrote the post. Databricks first showed the agent in June. It spent the months since in a private preview, watching real customer pipelines.
The number behind the beta
Databricks says the preview analysed "over 80,000 customer issues across hundreds of companies". The team writes that "over half of issues require organization-specific knowledge to properly triage, rank, and resolve".
That number drives the main new feature. Custom agent instructions let a team tell the agent what its own data means.
What the agent does, in five steps
The post names the five steps clearly. Detect a problem: a failed job, or a drop in data freshness. Investigate: read logs, source code, metadata and column lineage.
Verify a fix: run unit tests and a dry run in a sandbox. Notify the team through Slack, Teams or PagerDuty. Remediate by opening a pull request on a Git-backed job.
The last step is where Databricks has drawn a line. "ZeroOps will never change your code or data without your direct approval," the team writes. The agent proposes a fix. A person decides.
The agent reads what you already have. It uses Unity Catalog metadata, Delta history, Data Quality Monitoring signals and the Git repo behind a job.
The preview taught the team one thing about shape. A single failure usually has more than one cause. The agent needs to say so, not pick one and run.
New in the public beta
The beta adds a shared inbox so a team sees the same incidents. Notification summaries explain each alert in one line. One-click PR creation works for Git-backed jobs. Alerts redact PII.
The agent respects Unity Catalog permissions. It only reads what the person running it can read.
Databricks reports no success rate and no customer story in the post. Readers get the shape of the method, not a measured outcome.
A data team running Databricks pipelines can turn this on today. Every change stays behind a human review.
The 80,000-ticket number is the part worth asking about. Databricks does not say how many the agent fixed on its own. It does not say how many needed a human fix on top. It does not say how many it got wrong.
Ask for a run through a recent incident on your own pipeline before you decide. An agent that misreads column lineage can draft a confident pull request against the wrong table. The one-click PR button makes that fast.
Source
Genie ZeroOps, now in public beta, Databricks.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
