AI agents for Databricks
A Monmo agent works in Databricks the way a data engineer covering the morning shift would. It explains failed job runs, answers questions in SQL against Unity Catalog tables, tracks DBU spend from system tables and documents tables your team keeps asking about. It uses the identity you connect, sees only what Unity Catalog grants it, and asks before it runs a job, changes compute or writes data.
Updated 2026-11-13 · Free to start. Paid plans from $15 a month.
What agents do in Databricks
Explain a failed job run
When a scheduled job fails, the agent reads the run's task output, the error and the driver logs, finds the notebook cell or file that broke and writes the cause and a suggested fix.
AskThe nightly orders job failed at 02:10. Tell me which task broke, why, and what you would change.
Answer questions on a SQL warehouse
Ask for weekly active customers by segment and it finds the tables in Unity Catalog, writes the query, runs it on the SQL warehouse you assign and shows the answer with the SQL.
AskHow many weekly active customers did we have per segment over the last eight weeks? Show the query.
Track DBU spend
Every Monday it reads the billing usage system table, groups DBUs by workspace, job and SQL warehouse, compares them with the week before and writes the summary to Google Sheets.
AskEvery Monday, summarise last week's DBU use by job and warehouse, and point out the biggest increases.
Find compute that never sleeps
It lists all-purpose clusters without auto-termination, oversized clusters running light jobs and SQL warehouses that stay up overnight, and drafts the settings changes. Applying any of them asks first.
AskFind clusters without auto-termination and warehouses running all night, and suggest settings for each.
Document Unity Catalog tables
It reads a schema's tables, samples their columns and lineage, and drafts comments a new analyst could understand. You review them in Google Docs, and adding them to Unity Catalog waits for your approval.
AskDraft table and column comments for the sales schema in Unity Catalog. Let me review them first.
Compare experiment runs
For an MLflow experiment it lines up the runs by their parameters and metrics, picks out what changed between the best and the rest, and writes a short note for the team in Slack.
AskCompare the last 20 runs in the churn-model experiment and tell me which parameters made the difference.
Review who can see sensitive tables
It lists the grants on tables that hold personal data, the groups and users behind them, and any broad grant such as all account users, then drafts the revokes for your admin.
AskWho can read the tables in the customers_pii schema? Flag anything granted wider than the data team.
How it connects
- 1You connect a Databricks workspace with a personal access token or OAuth, or better, with a service principal your admin creates. Your workspace admin may need to allow tokens, and system tables must be enabled to report spend.
- 2Databricks fits a data agent and an engineering agent. Unity Catalog grants decide which catalogs and tables each one sees, so give the agent's identity only the schemas its work needs.
- 3Reading runs, logs, tables and system tables needs no approval. Running a job or notebook, creating or editing compute, writing to a table and changing a grant each ask first on a card.
Questions people ask
Can an AI agent work in Databricks notebooks?
Yes. A Monmo agent can read notebooks, explain what a cell does, and draft a new notebook for an analysis. Saving it to the workspace and running it on compute ask first, because a run uses DBUs and may write tables. It works through the Databricks connection, or in its own browser when the connection does not cover something.
Is it safe to give an AI agent access to Unity Catalog data?
The AI agent sees exactly what Unity Catalog grants its identity, no more. Give it a service principal with SELECT on the schemas it needs and nothing on personal data unless the job requires it. Reading needs no approval; any write or grant change waits for you, and you can remove access at any time.
Can it tell us why our Databricks bill went up?
It can show where it went up. From the billing system table it breaks DBUs down by job, cluster and warehouse, compares periods and names the jobs or clusters behind the increase. It suggests changes such as auto-termination or a smaller warehouse, and you decide which to apply.
Can it fix and rerun a failed Databricks job?
It finds the cause, drafts the fix and asks before rerunning. If the fix is in code, it opens a pull request in GitHub for review. If the job only needs repairing from the failed task, it shows you the run on a card and repairs it once you allow.