Databricks Expands Agent Platform

Databricks is doubling down on the agent revolution with a significant expansion of its Agent Bricks offering. Launched at the Data + AI Summit 2026, the platform aims to tackle the often-overlooked complexities of building and deploying AI agents.

The company states that while building the core agent loop is straightforward, the real challenge lies in the remaining 99% of development: managing token capacity, deployment, security, evaluation, monitoring, context, and sharing. Databricks observed developers bogged down in infrastructure work instead of agent creation.

The Agent Platform Imperative

To address this, Databricks is positioning Agent Bricks as a comprehensive platform designed to solve three critical challenges: Choice, Context, and Control.

Choice: Developers need access to a diverse range of models, from proprietary frontier models to smaller, cost-effective open-source options, and even custom-trained enterprise models. Databricks now offers native integration for models from OpenAI, Anthropic, Gemini, Qwen, Kimi, and notably, Grok models via a partnership with SpaceX.

Context: Agents require robust mechanisms to retrieve and process relevant data for accurate decision-making. This includes navigating messy data landscapes and unrecorded information. Innovations like agentic search, memory scaling, and programmable scratchpads are being delivered through components like Unity Catalog's MCP support for external data sources and the Genie Ontology.

Control: With agents accessing sensitive data, robust governance and cost management are paramount. The new Databricks agent platform introduces Unity AI Gateway, a unified governance layer for securing, observing, and managing AI assets. This includes fine-grained access controls, cost monitoring, and intelligent traffic routing.

Building Blocks for Agents

Databricks is integrating its research innovations directly into Agent Bricks. This includes a managed agent memory service, document intelligence features for parsing PDFs and other documents, and secure Databricks Sandboxes for isolated computation and data access.

The platform supports various agent harnesses like LangGraph and Agno, with horizontal autoscaling available via Databricks Apps. They are also offering a managed version of their open-source meta-harness, Omnigent.

Agent Traces and Monitoring are now integrated with LakeWatch, providing alerts for PII violations and auditing sensitive data access. Contextual Policies, written in SQL, allow for dynamic security rules based on data context.

Unity Catalog now registers agents, tools, and models, extending data governance principles to AI assets. This unification aims to provide consistent policies and end-to-end visibility across the data and AI lifecycle. This evolution positions the Databricks agent platform as a more complete solution for enterprise AI development.