Databricks Retires Supervisor API Beta for Custom Agents
Developers who used the managed agent-loop endpoint are being directed to Databricks Apps, where code, server behavior, deployment and identity choices become explicit parts of the application.
Edited by Tyronne Panaino
Databricks ended the Supervisor API beta on September 30, 2026 and says the endpoint is no longer available after that date. The retirement affects developers who used the API to send a model, tools and instructions in one request while Databricks handled the repeated agent loop, including model calls, tool selection, tool execution and synthesis of the final response.
The deprecated API documentation and the company's release-change record both direct users toward custom agents on Databricks Apps. That is a migration direction rather than evidence of a drop-in replacement: the Apps approach exposes more of the agent's code, server configuration, deployment process and identity model to the team operating it.
What the retired API managed
The Supervisor API was a programmatic route for building tool-calling agents through an OpenResponses-compatible endpoint. A request could select the model and tools at runtime while the platform managed the multi-turn loop. Databricks documented support for background execution and tools including Genie Agents, Unity Catalog functions, MCP servers, knowledge assistants, apps and serving endpoints.
That design reduced the amount of loop orchestration an application had to own. Its removal means teams should identify every production or test workflow that called the retired endpoint rather than assuming an unchanged request will continue to run. The official pages do not describe a compatibility period after September 30 or promise automatic translation of an existing Supervisor request into an Apps deployment.
Databricks Apps is the recommended destination
The custom-agent guide describes an application-based architecture in which developers control the agent code, server configuration and deployment workflow. Databricks provides templates that combine an agent, a conversational API, a chat interface and MLflow evaluation scaffolding, but teams can also begin from their own local project.
Databricks recommends the MLflow `ResponsesAgent` interface for connecting custom frameworks to platform features. The guide says an existing agent can be wrapped for compatibility with AI Playground, Agent Evaluation and Agent Monitoring, with tracing available through MLflow. That can preserve some managed platform services while leaving the application responsible for its own server and deployment lifecycle.
The migration therefore changes the boundary of responsibility. Under the retired API, Databricks ran the agent loop from one request. Under the Apps path, a team packages and operates an application that contains the loop or framework it chooses. The practical work is not only replacing an endpoint; it includes validating tool behavior, deployment automation, observability and failure handling in the new application shape.
Identity needs an explicit decision
The Apps guide documents two authorization models. App authorization uses a service principal created for the application, so users share the application's permissions. User authorization acts on behalf of each person and applies that person's individual access. A migration should choose between those models deliberately because the choice affects which data and tools an agent can reach and how actions are attributed.
This run did not execute a Supervisor workload or test an Apps migration. The evidence establishes the end-of-life date, the former API's role and Databricks' recommended destination, but not migration effort, behavioral parity or reliability for a specific agent. Teams should validate an affected workflow end to end before considering the transition complete.
Status
Confirmed. Databricks says the Supervisor API beta reached end of life on September 30 and is no longer available. Internal confidence is medium because the migration guidance comes from Databricks and was not independently exercised in this run.
Sources
- Databricks — Supervisor API deprecated documentation
- Databricks — Upcoming release and behavior changes
- Databricks — Author and deploy a custom agent on Databricks Apps
Update note: Last reviewed 2026-10-06. We will revise this post if Databricks publishes additional migration tooling, compatibility guidance or independently verifiable transition evidence.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.