topic Re: Testing out Agentic Capabilities in Generative AI

Testing out Agentic Capabilities

Tinjar — Wed, 15 Oct 2025 06:13:45 GMT

So I am creating a POV on Databricks' agentic capabilities and wanted to showcase its abilities through a simple change pipeline.

A user asks for changes in a specific table in a schema -> based on metadata info from our lake table info is received -> the sql for this table is fetched from a repository such as github/bitbucket -> changes are done to the sql and tested -> The modified sql is then pushed back to the repo.

The approach I am currently thinking of is to do it through the data bricks assistant Data Science agent and providing it python functions as tool calls and allowing it to call the functions from the notebook for each of these steps.

My question is, is this viable in the first place? Also, is this the best way of tackling this use case solely through Databricks' in house agents. To give you context, the other povs we are testing for are similar coding agents such as Codex

Re: Testing out Agentic Capabilities

AbhaySingh — Fri, 17 Oct 2025 02:17:41 GMT

Yes, your approach seems fairly viable . Here are some thoughts.

Step-by-Step Viability

User requests - Supported via Mosaic AI Agent Framework + ResponsesAgent interface
Metadata retrieval - Unity Catalog Functions can query INFORMATION_SCHEMA tables directly
Fetch SQL from repo - Databricks Repos API + Databricks SDK in Python UC Functions
Modify & test SQL - LLM-powered modification + SQL Statement Execution API for testing
Push to repo

Re: Testing out Agentic Capabilities

Tinjar — Wed, 29 Oct 2025 11:21:56 GMT

Thanks a lot Abhay!

Re: Testing out Agentic Capabilities

CharlotteMarti2 — Wed, 29 Oct 2025 12:29:31 GMT

Yes, it’s viable — Databricks’ in-house agents can handle that workflow if you define each stage as callable Python tools. The key is robust function design and proper metadata access. However, you might find integrating Unity Catalog and Git integration APIs simplifies repo updates and version control.