Databricks Genie ZeroOps public beta: a pipeline agent that files PRs
Databricks Genie ZeroOps is in public beta: an agent that watches jobs and tables, finds the root cause and proposes a fix as a PR. Protect your main branch first.
The worst data bug is the job that succeeds and writes the wrong numbers. On November 19, Databricks put Genie ZeroOps in public beta. It is a background agent that watches your jobs, pipelines and tables, finds what broke and why, and proposes a fix. For Git-backed jobs, it can open the pull request for you.
What actually happened
From the Databricks announcement and docs:
- What it watches. Failing jobs and pipelines, jobs that succeed but leave stale or incomplete tables, and tables that go stale. Root causes include code changes, schema changes and upstream failures.
- The preview numbers. Databricks says ZeroOps saw more than 80,000 real customer issues across hundreds of companies in private preview, and that over half needed company-specific knowledge to triage.
- New in the beta. A shared inbox, summary notifications to Slack, Teams, PagerDuty or webhooks (no email), custom instructions in a markdown file, and one-click PR creation.
- Approval. The docs say it "doesn't change production data or code unless you approve the change." It never writes to your repo in the background.
- The trap. If you name your default branch as the target, the docs warn it commits straight to that branch and skips PR review.
- Limits. Workspace level only. Not HIPAA compliant. No models or apps, no cost or performance regressions. Table data-quality checks need separately billed data quality monitoring. A workspace admin turns it on from the Previews page.
The original June post said ZeroOps would cover ML models too. The beta docs say models are not monitored. We go with the docs.
Why it matters for your business
If your reporting runs on a warehouse, a silent bad table costs more than a loud failed job. Someone makes a pricing or inventory call on stale numbers. An agent that flags "this succeeded but the table is wrong" is worth having.
The useful part is the custom instructions file. Over half the preview issues needed local knowledge: which source is always flaky, which fix you prefer. Write that down and the agent gets better. Leave it blank and you get generic advice.
Before you turn it on:
- Turn on branch protection on your default branch. Then the agent can't skip review, even with a bad config.
- Write the instructions file with your known flaky sources and naming rules.
- Check what notifications send. Summaries leave Databricks and sit in Slack under Slack's retention, not Databricks'.
Key takeaways
- Genie ZeroOps entered public beta on November 19
- It finds failed jobs and silently stale tables, then proposes a fix or a PR
- Nothing changes in production without approval, unless you target the default branch
- Workspace-only, not HIPAA compliant, and no model monitoring in the beta
- Protect main and write the custom instructions before you enable it
Do your reports run on numbers nobody checks? We build data pipelines with freshness checks, alerting and protected branches, so a bad table gets caught before it reaches a decision. Tell us where your numbers come from.
Sources: Databricks blog: Genie ZeroOps, now in public beta, Databricks docs: Genie ZeroOps, Databricks docs: Get started.
- #databricks
- #genie-zeroops
- #data-pipelines
- #ai-agents
- #devops
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
OpenRouter drops 3 providers, pulls ZDR from 125 endpoints
OpenRouter off-boarded three inference providers and removed 125 endpoints from zero data retention after Dragos canary research. What to check in your stack.
Read itOpenAI Performance Factory: Codex agents debug, engineers merge
OpenAI's Performance Factory ran 1,700+ Codex investigations into ChatGPT slowdowns. 70% of audited root causes held up; about 15 agent PRs shipped as written.
Read it