supervisor-api

Call the Supervisor API via `responses.create()` after declaring hosted tools as a structured list to delegate iterative tool-selection, looping, and synthesis logic to Databricks server-side.

Yes, you can mix Databricks-hosted tools like Genie spaces, Unity Catalog functions, and knowledge assistants with your own Python-callable client-side function tools within the same `responses.create()` request.

The Databricks Supervisor API is used to replace client-managed agent tool looping with server-side orchestration. It handles the iterative execution and final synthesis of AI responses, removing the burden of managing tool-selection logic on the client.

No, the Supervisor API enforces mutually exclusive stream and background modes. You must choose one mode per request and follow constraints around unsupported inference parameters and token scoping, as on-behalf-of (OBO) token scoping is not supported.

To enable MLflow distributed tracing for server-side agent loops, apply the optional MLflow tracing configurations alongside drop-in changes for `agent_server/agent.py`, which supports model selection at runtime while calling `responses.create()`.