Databricks gives Genie Agents a longer analytical loop — and an API

Databricks has expanded Genie Agents with an agentic analysis mode, external APIs, governed file reasoning and new curation tooling — a bundle that pushes its natural-language analytics product beyond one-shot SQL generation toward longer-running analytical work.

From a query to a research plan

The headline change is Agent mode, which Databricks says is now available for all Genie Agents. Instead of stopping after one query, the mode can create and refine a research plan, run multiple iterative queries, and return a report containing findings, visualizations and citations. Databricks positions Chat mode for quick factual lookups and Agent mode for more open-ended investigations.

That distinction matters for NL2SQL deployments. A production analytical question often requires several dependent steps: establish the relevant population, inspect changes over time, test possible explanations and then present the evidence. The new loop makes those stages part of the product rather than something an application team must orchestrate around a single generated query.

Databricks is also exposing Agent mode through APIs. According to the release, developers can embed the workflow in custom applications, chatbots, scheduled reports and internal tools; responses stream through Server-Sent Events, while conversations and messages can be retrieved for monitoring. The API also supports downloading visualizations and data tables attached to responses.

Structured and unstructured context in one conversation

Genie Agents can now analyze files held in Unity Catalog volumes alongside governed tables. Builders may attach up to 10 volumes containing formats such as PDFs, documents, slide decks and images. Databricks says file analysis honors Unity Catalog permissions and can use content-search indexing to improve retrieval across larger collections.

For practitioners, that creates a useful but demanding deployment pattern: an agent can connect a metric in a warehouse to context in surveys, contracts or operating documents, while the platform keeps access tied to the same governance layer. Evaluation now has to cover both query correctness and whether the agent selected and interpreted the right supporting documents.

Curation becomes an operating loop

The update also extends Genie Code, Databricks’ assistant for building Genie Agents. It can generate a baseline agent from a purpose, example questions and data sources; review conversation failures or benchmark runs; and summarize recurring user topics and feedback. Authors remain responsible for reviewing suggestions before saving instructions, example SQL or knowledge-store configurations.

That human approval point is important. The release does not eliminate semantic curation; it tries to shorten the cycle between observing a bad answer and improving the context that produced it. The practical shift is from “generate SQL from a prompt” to operating a domain-specific analytical agent whose plans, retrieval, queries and outputs all need monitoring.