databricks-model-serving

Register the model to Unity Catalog with mlflow autolog or log_model, then create an endpoint using the MLflow deployments SDK, Databricks SDK, or the Serving UI. Specify the entity name, version, workload size, and scale-to-zero settings.

Subclass MLflow 3's ResponsesAgent, implement predict and predict_stream methods, and use ChatDatabricks with a Foundation Model endpoint. Export the agent with mlflow.models.set_model and log it with mlflow.pyfunc.log_model referencing the agent file.

ResponsesAgent requires helper methods instead of raw dictionaries for output items. Use self.create_text_output_item(text, id) for text responses, and the function call helpers for tool calls and results.

Use UCFunctionToolkit from databricks-langchain with function names like catalog.schema.function, then bind the tools to your LLM. Register DatabricksFunction resources when logging the model so authentication passes through automatically.

Agent deployment takes about 15 minutes, exceeding synchronous tool timeouts. Use a job-based approach: create a deployment job with manage_jobs, trigger it with run_now, and poll status with get_serving_endpoint_status.

DBR 16.1 or later is recommended because it pre-installs MLflow 3, LangChain, and pydantic. You still need to pip install databricks-langchain, langgraph, and databricks-agents for agent development.