Overview
At Data + AI Summit 2026, Databricks announced the expansion of Agent Bricks into a comprehensive enterprise agent platform. The central premise: the basic agent loop represents only 1% of real work; the remaining 99% is hidden infrastructure — token capacity, deployment, security, evaluation, monitoring, context management, and team collaboration.
Over 100,000 agents have been built on the platform since launch, with customers including AstraZeneca, 7-Eleven, Fox Corporation, and Block. The platform is organized around three pillars: Choice, Context, and Control.
Three Pillars of Agent Bricks
1. Choice — Models and Frameworks
Databricks supports frontier proprietary models (OpenAI, Anthropic, Gemini, Qwen, Kimi) and open-source alternatives. A major announcement is the partnership with SpaceX to provide Grok models natively on the platform.
Custom model training capabilities include prompt optimization, fine-tuning, and reinforcement learning. A Databricks-trained custom data agent demonstrated competitiveness with Anthropic’s Opus and Sonnet models while significantly reducing per-query costs.
Supported agent frameworks include LangGraph, Agno, CrewAI, Claude Code SDK, OpenAI Agent SDKs, and the Omnigent meta-harness now available as a managed service. Horizontal scaling is handled through Databricks Apps.
2. Context — Data Access and Advanced Retrieval
MCP with Unity Catalog enables secure connections to external sources like Google Drive, JIRA, Slack, and GitHub with specialized search capabilities. Genie Ontology continuously learns business semantics: fiscal calendars, department heads, corporate definitions, data lineage, and author authority.
Document Intelligence (now GA) includes SQL functions like ai_parse_document, ai_extract, and ai_classify for PDF and document analysis. The Agent Memory Service, powered by Lakebase, provides persistent memory for session history and cross-agent coordination.
The Databricks Agent Tools suite includes a document search subagent reported as “3x faster than before, while improving quality.” Databricks Sandbox provides secure VMs with downscoped Unity Catalog access for code interpreters and experimentation.
3. Control — Governance, Security, and Cost Management
Unity AI Gateway acts as a unified governance layer with:
- Catalog discovery of agents, models, MCPs, Skills, and external agents
- Fine-grained access controls for tools and agents
- Per-user and per-group budget enforcement with hard spend caps
- Intelligent traffic routing based on reliability and policy compliance
Agent Traces and Monitoring: reasoning traces, memory writes, and generations are stored in the Lakehouse for full analytics and debugging. Lakewatch integration detects PII violations and audits sensitive data access.
Contextual Policies (SQL-based, Python support coming) enable stateful security controls based on data context. For example: restricting PII-containing data publication to websites while permitting email sharing, or requiring human approval for specific actions like Salesforce updates.
Key Points
- 100,000+ agents built on the platform since launch
- Support for all frontier models: OpenAI, Anthropic, Gemini, Grok (SpaceX), Qwen, Kimi
- Omnigent available as a managed service to orchestrate existing agent workflows
- Agent Memory Service with Lakebase for persistent memory across sessions
- Document Intelligence GA with SQL functions to process PDFs directly in queries
- Unity AI Gateway with hard spend caps and intelligent routing
- Contextual SQL policies for stateful security governance
- Databricks Sandbox for secure code execution and experimentation
- Lakewatch integration for real-time PII violation detection
Why It Matters
Agent Bricks represents the maturation of the agentic AI ecosystem. Until now, companies could build agents relatively quickly, but production deployment was a costly, uncertain process: how do they scale? how are they audited? how do you govern costs when an agent starts querying models autonomously?
Databricks’ proposition is that governing data and governing AI on the same platform eliminates the structural friction that comes from using disconnected tools. An organization that already has Unity Catalog for its data can extend exactly the same policies to its agents, using the same SQL language their data teams already know.
The combination of model choice, sophisticated context retrieval, and fine-grained control addresses the practical barriers that have prevented widespread agentic application deployment in enterprises — making production-grade AI agents accessible beyond the largest tech companies.
