Skip to main content

Agent Platform

The Databricks Agent Platform provides the building blocks for creating, deploying, and governing custom agents. The platform spans five areas — Framework (build with any harness and any framework), Model (access any model through Unity Gateway), Compute (Agent Runtime and Agent Sandbox), Context (Agent Memory and Unity Catalog), and Observability (tracing and evaluation through Unity Gateway and MLflow). For connected partners, the Agent Platform is how you build custom agents on Databricks — either for your own product or for customer use cases. These agents are then callable through the OpenAI-compatible API as described in Integrating with Agents.

Agent Bricks​

Agent Bricks is the Databricks platform for building and deploying custom agents. Partners develop agents using any framework (LangChain, LangGraph, LlamaIndex, DSPy, CrewAI) and any model — foundation models served through Unity Gateway or external providers registered as Unity Catalog securables. Agents are logged and versioned with MLflow, deployed through Agent Runtime, and governed by Unity Catalog throughout.

Custom agents built with Agent Bricks are called through the OpenAI-compatible API — using the Databricks OpenAI Client (recommended) or direct REST calls via the Responses or Chat Completions interface. See Integrating with Agents — Agent Bricks for calling patterns.

Documentation: Agent Bricks | Author AI agents in code | MLflow 3 for GenAI

Agent Runtime​

Agent Runtime is the compute layer that hosts and deploys custom agents. Partners deploy agents to Databricks Apps. Agent Runtime provides production-grade reliability, latency management, and enterprise-grade security and compliance.

Documentation: Deploy an agent for generative AI application | Model Serving | Databricks Apps

Agent Memory​

Agent Memory provides short- and long-term memory capabilities for agents on Databricks, enabling agents to retain and recall context across interactions. Databricks offers two approaches: managed agent memory (Beta) — Databricks manages the infrastructure, governed by Unity Catalog — and self-managed agent memory using Lakebase for direct SQL access to memory state.

Documentation: Agent memory | Managed agent memory | Self-managed agent memory (Lakebase) | Memory API reference

Agent Sandbox​

Agent Sandbox provides isolated, secure sandboxes for agents to execute generated code safely within the customer's workspace. Sandboxes spin up in seconds, persist storage across sessions, and support concurrent SSH sessions. Partners can use the system.ai.python_exec Unity Catalog function (exposed as an MCP tool) to let agents run dynamically generated Python in an isolated environment with no internet access and scoped Databricks data access only.

Documentation: Databricks Sandbox | Let agents run code

What's next​