Sep 25 — Oct 2, 2026
Agentic AI infrastructure dominated the ecosystem this week with the general availability of Unity AI Gateway and Lakebase Search. Core developer workflows also saw key quality-of-life upgrades across Databricks Asset Bundles and the introduction of the ai_decide inference function.
1.Lakebase Postgres rolls out native vector search and branch-based instant recovery
Lakebase Postgres announced the general availability of Lakebase Search on AWS and Azure, delivering native full-text, vector, and hybrid search alongside operational data without separate ETL pipelines. In parallel, Databricks introduced branch-based restores that execute instant, copy-free recoveries via decoupled storage, while developer tooling added branch expiration controls and parallel agent sandbox workflows.
Sources
- Lakebase Postgres branch-based restores for fast recovery at scaleNews · databricks-blog · Oct 1
- Parallel Coding Agents with Lakebase | Claude Code + GitHub ActionsVideo · Databricks · Oct 1
- Databricks Lakebase Search - GA AnnouncementsCommunity · reddit · Sep 30
- Lakebase Search: State-of-the-art full text and vector search for PostgresNews · databricks-blog · Sep 28
2.Unity AI Gateway formalizes multi-model governance and MCP agent execution
Databricks made Unity AI Gateway generally available, establishing centralized multi-model governance, cost tracking, and tracing for production agents. Updates across the Go and Java SDKs introduced new Unity Catalog securables for agent skills, while practitioners gained the Unity Gateway CLI and reference architectures for running Model Context Protocol (MCP) servers alongside Genie agents.
3.ai_decide introduces low-cost System One routing directly to SQL and SDKs
Databricks unveiled ai_decide, a specialized function designed for fast, deterministic classification and structured decisions directly within data pipelines rather than token-by-token text generation. SDK support landed in Python and Java, with practitioners quickly adopting the primitive as a lightweight upstream router to triage queries before invoking expensive foundation models.
Sources
- The Shift to System One | System One Models - Part OneVideo · Advancing Analytics · Oct 2
- Routing between GLM 5.3 Flash and GLM 5.3 with ai_decide: a cheap router in front of ai_queryCommunity · databricks-community · Oct 1
- Introducing ai_decide: make fast decisions on your governed dataNews · databricks-blog · Sep 30
- v0.144.0Release · databricks/databricks-sdk-py · Sep 30
4.Databricks Asset Bundles and Jobs add on-deploy triggers and structured environment variables
Orchestration workflows saw tighter deployment integration, highlighted by support for running jobs automatically upon bundle deployment via on_bundle_deploy. Simultaneously, breaking SDK changes standardized required environment variable definitions across Jobs and Tasks, complemented by direct engine deployment migrations in the VS Code extension.