Databricks Launches Mosaic Agent Bricks: Automating AI Agent Optimization for Enterprise Production

On June 11, 2025, at the Data + AI Summit, Databricks introduced Mosaic Agent Bricks, a new platform designed to address the common issue of enterprise AI agents failing to reach production. Despite advanced AI technology, many companies struggle to deploy these agents effectively due to inefficient and manual evaluation processes. Databricks aims to streamline and automate these processes to enable better, faster, and more reliable AI agent deployment. The root of the problem, according to Hanlin Tang, Databricks' Chief Technology Officer of Neural Networks, lies in the lack of robust evaluation mechanisms. Most enterprises rely on manual assessments, which are time-consuming, inconsistent, and scalability issues. Tang emphasized that this approach often results in "quality drift," cost overruns, and missed deadlines, making it difficult for companies to gain the confidence needed for production deployment. Mosaic Agent Bricks builds upon the Mosaic AI Agent Framework, which Databricks acquired in 2023 for $1.3 billion. Prior to the acquisition, Mosaic focused primarily on research, which saw limited immediate enterprise impact. However, once integrated with Databricks, the team gained direct access to thousands of enterprise customers, revealing significant pain points and new research opportunities. This led to the development of Mosaic Agent Bricks, which is designed to automate the entire AI agent optimization pipeline. Key innovations in Mosaic Agent Bricks include Test-time Adaptive Optimization (TAO), a method that tunes AI models without labeled data; domain-specific synthetic data generation; task-aware benchmarks; and an automated quality-to-cost balance optimizer. These features collectively aim to make the development and deployment of AI agents more manageable and effective. The process begins with a high-level task description provided by the user and the enterprise's data. The platform then generates task-specific evaluations and judges, creates synthetic data mirroring the customer's dataset, and searches across various optimization techniques to find the most effective configuration. Tang highlighted that this approach eliminates the need for detailed technical specifications, making the platform more accessible and user-friendly. Mosaic Agent Bricks is not an isolated solution but is part of Databricks' broader unified data stack. The company announced the general availability of its Lakeflow data engineering platform, which unifies data ingestion, transformation, and orchestration—previously requiring separate tools. Lakeflow streamlines the process of preparing data for AI tasks, ensuring that the data is clean, structured, and efficiently managed. This integration ensures that Mosaic Agent Bricks can operate on well-prepared data, enhancing overall performance and reliability. Further, Databricks' Unity Catalog provides robust data governance features, including access controls and data lineage tracking. This integration ensures that the behavior of AI agents respects enterprise data policies, reducing compliance risks and enhancing trust. One of the standout features of Mosaic Agent Bricks is the Agent Learning from Human Feedback (ALHF) mechanism. ALHF allows for natural language guidance to adjust multiple system components, addressing the issue of "prompt stuffing," where users over-pack prompts with guidance, leading to suboptimal results. Tang explained that ALHF interprets vague human instructions and automatically adjusts the relevant components, making the agents more steerable and responsive to user needs. Mosaic Agent Bricks supports two primary methods for agent-to-agent communication: exposing agents as endpoints that can be wrapped in various protocols and using a multi-agent supervisor that is Model Context Protocol (MCP) aware. This flexibility ensures compatibility with a wide range of existing and emerging standards. The strategic implications for enterprise decision-makers are significant. Deploying AI agents without proper evaluation mechanisms can lead to suboptimal outcomes, and lacking a solid data foundation can further exacerbate these issues. Mosaic Agent Bricks addresses this by providing a comprehensive, automated evaluation and optimization framework. This enables organizations to focus on identifying valuable use cases and preparing their data, rather than spending extensive resources on building and tuning evaluation frameworks. Industry insiders view Mosaic Agent Bricks as a game-changer in the agentic AI landscape. The platform's automation and optimization capabilities, combined with Databricks' robust data stack, position it to significantly reduce the barriers to effective AI agent deployment. For enterprises aiming to lead in AI adoption, this development means they can accelerate their AI initiatives without compromising on quality or governance. Databricks, known for its unified data and AI platform, continues to innovate and deliver solutions that cater to the unique needs of enterprise environments.