Databricks announces expansions to its Agent Bricks framework.
Databricks significantly expands its Agent Bricks framework with new features around governance, evaluation, and data access. This way, the company aims to deploy its AI agents safely and reliably in production. During its “Week of AI Agents,” Databricks announced updates for MLflow, a proprietary governance layer for models, and a marketplace for MCP servers.
Evaluation Gets an Upgrade
MLflow, originally focused on machine learning, now also supports the evaluation and monitoring of AI agents. Organizations can train their own judges with domain-specific criteria. Feedback in natural language is used to adjust these judges, both on test data and live traffic. This is important in environments where reliability matters, including checks for bias, robustness, and decision quality.
The AI Gateway is new, a governance layer on top of both commercial and open-source models. This allows Databricks to enforce logging, access control, and audits on endpoints from GPT-5, Gemini, Claude, or Llama, for example. All of this is managed through Unity Catalog, with options to shield usage costs.
Marketplace for External Context
With support for the Model Context Protocol (MCP), AI agents can safely call third-party data and services. The online store grows based on customer demand. Partners can easily connect, though some services incur usage costs. A new Multi-Agent Supervisor (in beta) can coordinate workflows across multiple agents and MCP servers, with automated actions such as support tickets or SQL queries.
The new capabilities are available starting today, with some components in public preview or beta, writes SiliconANGLE.