Databricks: How AI agents are intended to make the leap to productive operation

Databricks wants to make it easier for companies to move from AI prototypes to productive applications. At the Data+AI World Tour in Munich, the company presented new tools designed to make AI agents safer, more controllable, and more economical to operate in the future.

The focus was on the new Agent Bricks platform, several extensions to the Data Intelligence Platform, and direct integration with SAP Business Data Cloud. This allows SAP data to be integrated live and bidirectionally into Databricks via Delta Sharing—without replication. Companies can thus combine financial, production, or supply chain data directly with external sources, for example, for forecasts or risk analyses.

Agent Bricks: AI Agents in Series

According to Databricks, the reason companies fail when transitioning from AI prototypes to regular operation is not so much a technical problem as it is usually costs, lack of governance, and data quality. The company aims to address exactly these points with its new platform strategy.

Agent Bricks is intended to standardize the development and productive operation of AI agents. The platform offers a unified environment for creating, evaluating, and orchestrating agents—regardless of model or framework. Three approaches are intended to appeal to different user groups:

• Agentic AI Functions: For data engineers who want to integrate AI-powered transformations such as document or log analyses directly into existing data pipelines.
• Declarative Agents: A low-code approach where users can define and test agents using natural language and templates.
• Custom-Code Agents: For developers who want to use frameworks like LangChain or models like GPT-5, Claude, Llama, or Gemini on Databricks' infrastructure.

The innovations presented in Munich also primarily aim at quality, governance, and orchestration:

• ai_parse_document: A function that extracts structured information from PDFs and unstructured texts. This allows, for example, contract contents or technical reports to be analyzed automatically.
• MLflow 3.0 for Agent Observability: extends the open-source tool MLflow with functions for observing and evaluating agents. Every interaction is logged to make performance and quality traceable.
• AI Gateway and MCP Catalog: Provide control and auditability for external models and tools. Companies can manage, limit, and bill API calls centrally—an important step towards cost and compliance transparency.
• Multi-Agent Supervisor: Enables the coordination of multiple specialized agents to automate complex business processes.

Data Quality at the Core

In Munich, Databricks once again emphasized its strategic line: data first, then intelligence. The Data Intelligence Platform, an evolution of the Lakehouse concept, is intended to create a robust foundation for trustworthy AI. The Unity Catalog, which manages data sources, access rights, and lineage, is central to this.

“We are talking about governance by design—data management should not be a subsequent patchwork,” says Databricks manager Matthias Ingerfeld. With Agent Bricks and the SAP integration, Databricks underscores this approach: the success of enterprise AI depends less on the size of a model than on a stable, controlled infrastructure.

While providers like Microsoft and SAP are leveraging their strengths in specific application scenarios and OpenAI is focusing on the further development of large language models, Databricks is taking a different approach: an open, vendor-neutral data and governance architecture. With competitors like Snowflake, ServiceNow, and Microsoft (Azure Fabric), Databricks is thus competing for the same customer group—companies that want to systematically and scalably integrate AI into their data landscapes. Whether this focus will lead to a breakthrough in practice remains to be seen.

(olb)