First of all, great set of tools and skills, thanks for the work on compiling these. Combined with recent development on the databricks CLI, working with databricks from terminal is super easy.
One element of the toolkit which is not working well is the framing that any questions about data go via databricks genie ask, coming from the databricks-data-discovery skill:
This skill **routes data work** — decide first:
- the request is about *the data* — finding it, asking questions of it, or
generating SQL → delegate to **Genie One**:
`databricks genie ask -s <session-label> "..."` (see Routing below).
I think the skill needs to be invoked ONLY when the agent doesn't have enough information to run queries on it's own. Currently very basic tasks get routed to it, which makes the agent slower and waste tokens. I was asking about describing the tables in a specific schema and I got this via codex and gpt-5.6-sol high:
databricks genie ask -s table-pairs "In catalog.schema, list every current table
or view and its object type. For these six specifically if present .... — return
every column in ordinal order with data type and the table comment." --include-sql
This wastes time and tokens. Before adding this skill, the model would write an SQL query to get back the information schema of the catalog, but now I'm doing an extra loop through Genie.
There are genuine use cases for the skill, where Genie sees more of the metadata (finding data where you don't know where it lives, frequent joins, queries), and this skill works great for those cases. But I think routing every single question through it is a bit pointless considering most questions require nothing more than databricks experimental aitools tools query ..., which is both cheaper and faster.
EDIT: This is also causing agents to do worse on data work when tables don't need to be discovered. Telling the agent to stop asking Genie and do the work itself made the outcome better and faster, but I don't want to have to disable the data discovery skill for when it's useful.
First of all, great set of tools and skills, thanks for the work on compiling these. Combined with recent development on the databricks CLI, working with databricks from terminal is super easy.
One element of the toolkit which is not working well is the framing that any questions about data go via
databricks genie ask, coming from thedatabricks-data-discoveryskill:I think the skill needs to be invoked ONLY when the agent doesn't have enough information to run queries on it's own. Currently very basic tasks get routed to it, which makes the agent slower and waste tokens. I was asking about describing the tables in a specific schema and I got this via
codexandgpt-5.6-sol high:This wastes time and tokens. Before adding this skill, the model would write an SQL query to get back the information schema of the catalog, but now I'm doing an extra loop through Genie.
There are genuine use cases for the skill, where Genie sees more of the metadata (finding data where you don't know where it lives, frequent joins, queries), and this skill works great for those cases. But I think routing every single question through it is a bit pointless considering most questions require nothing more than
databricks experimental aitools tools query ..., which is both cheaper and faster.EDIT: This is also causing agents to do worse on data work when tables don't need to be discovered. Telling the agent to stop asking Genie and do the work itself made the outcome better and faster, but I don't want to have to disable the data discovery skill for when it's useful.