Databricks Mosaic AI Agent Framework

Databricks Mosaic AI Agent Framework is an enterprise platform for building, evaluating, and deploying production AI agents with deep integration into the Databricks Lakehouse Platform. It provides a comprehensive environment where data teams can build agents that leverage their organization's data assets directly, without complex data pipeline engineering.

The Agent Framework includes tools for building RAG agents with automatic retrieval from Unity Catalog-governed data sources, including Delta tables, vector search indexes, and unstructured documents. Agents can query structured data through natural language, access feature stores, and leverage ML models registered in MLflow — all within the governance framework of Unity Catalog.

A standout capability is Mosaic AI Agent Evaluation, which provides systematic testing of agent quality with LLM-as-judge scoring, retrieval accuracy metrics, and custom evaluation criteria. The evaluation framework integrates with MLflow Experiments for tracking agent performance over time and comparing different agent configurations.

Databricks Model Serving provides the deployment infrastructure, offering scalable endpoints with built-in monitoring, A/B testing, and automatic scaling. Agents can be served alongside the data they need, eliminating the latency and complexity of external data access.

The platform supports LangChain, LlamaIndex, and custom Python agents, with hosted access to foundation models including DBRX, Llama, and Mixtral through Databricks Foundation Model APIs. Pay-per-token pricing on foundation models and serverless compute for agent serving make costs predictable. For enterprises already using Databricks for data and ML, the Agent Framework provides a natural path to production AI agents with enterprise governance.