Databricks Supervisor API dies 30 September 2026, and it is not in the August release notes

Summary. Databricks has deprecated the Supervisor API and set its end of life at 30 September 2026, roughly five weeks from 23 August 2026. Both documentation pages carrying the notice were updated on 21 August 2026. Neither the August 2026 platform release notes on docs.databricks.com nor the Azure Databricks mirror of the same month mentions the Supervisor API at all: across 66 entries covering 4 model additions, 11 connector changes and 4 separate ABAC entries, the one item with a hard shutdown date is absent. A team that reviews monthly release notes for breaking changes will not see this one. Worse, Databricks' own preview release policy says Beta features are "intended to advance to either Public Preview or GA" and that Beta is on by default for Premium customers, so some workspaces have the feature enabled without anyone having asked for it.

What actually changed

The Supervisor API is a managed agent loop. You define the model, the tools and the instructions in one request to an OpenResponses-compatible endpoint at POST ai-gateway/mlflow/v1/responses, and Databricks runs the loop for you: calling the model repeatedly, selecting and executing tools, and synthesising the final response. It shipped as a Beta gated behind the workspace Previews page and required AI governance with Unity AI Gateway plus OpenTelemetry trace storage in Unity Catalog.

Two documentation pages now carry a deprecation warning, both stamped 21 August 2026. The Supervisor API reference reads: "The Supervisor API is deprecated and reaches end of life on September 30, 2026. After that date, it will no longer be available. Databricks recommends migrating to custom agents on Databricks Apps." The tutorial page for building an app on top of it carries the same date but a different instruction: "To keep building custom agents on Databricks Apps, use a custom agent loop instead."

Those two sentences are not the same migration. The first reads as a platform move. The second is accurate: the replacement is code you write.

The part that will catch people

Three details decide whether this is a footnote or a five-week project.

The first is the naming. Supervisor API is not Supervisor Agent, the Agent Bricks product that coordinates Genie Agents, agent endpoints, Unity Catalog functions and MCP servers with human feedback optimisation. Supervisor Agent is not deprecated. The two products sit three clicks apart in the same documentation tree and share a word. Expect at least one engineer on the thread to check the wrong page and report that nothing is wrong.

The second is default-on exposure. Databricks' release-types table states that for Premium customers Beta features are on by default, and for Enterprise customers they are off by default. The previews management page confirms previews are managed at account and workspace scope. A Premium workspace can therefore have had the Supervisor API available without a deliberate opt-in, which is exactly the condition under which a proof of concept quietly becomes a dependency.

The third is that the authoring guide still points at it. The Databricks Apps agent authoring page, last updated 3 August 2026, carries a tip: if your agent uses only Databricks-hosted tools and needs no custom logic between tool calls, use the Supervisor API to let Databricks manage the agent loop. The label reads "(deprecated)" inline, but the recommendation itself has not been rewritten. Anyone following the current tutorial is being pointed at a feature with 5 weeks left.

What Databricks' own policy says about Beta

The release-types page is worth reading before anyone argues this was foreseeable.

Release type	For production?	Interface stable	SLA	Stated trajectory
Private Preview	No	No	No	Invite only, often undocumented
Beta	No	No	No	"intended to advance to either Public Preview or GA"
Public Preview	Yes	Yes	Yes	Stable and intended to advance to GA
General Availability	Yes	Yes	Yes	Fully supported and production-ready

Beta carries no production recommendation, no interface stability and no SLA, with support routed to the engineering team rather than the support team. On paper, nobody should have shipped against it. In practice, "on by default for Premium" plus an OpenAI-shaped endpoint plus a documented tutorial is how Beta code reaches production. The stated trajectory for Beta is promotion, not retirement, which is why a hard shutdown date deserved a release-note line and did not get one.

How to tell whether this is you

Check for the endpoint first. Any code that posts to ai-gateway/mlflow/v1/responses is on the deprecated path. Then check whether a Databricks App was built from the Supervisor API tutorial rather than the plain custom-agent template; the tutorial's agent.py makes a single API call instead of running a loop, which makes it easy to spot. Finally, check the Previews page at both account and workspace scope for the Supervisor API entry, and check the Unity AI Gateway inference tables for traffic against that endpoint.

Agent work that already runs its own loop is unaffected. So is anything built on Agent Bricks Supervisor Agent, on Genie Agents, or on the AI Gateway unified and native APIs, which Databricks lists as the third build approach and which is not touched by this notice.

What the migration actually costs

The replacement is a custom agent loop inside a Databricks App. Databricks Apps run on the serverless platform, are billed per hour of compute while running, and are capped by a per-workspace resource limit. That is a different cost shape from a per-request managed endpoint, and it needs a line in the budget rather than an assumption of parity.

The engineering cost is the loop itself: model call, tool selection, tool execution, response synthesis, plus error handling and streaming. The upside is that the tutorial page already documents what you get back by writing it yourself. Databricks lists five cases where the Supervisor API could not be used and a custom loop was required: client-side function tools, which the Supervisor API cannot mix with hosted tools in one request; agent endpoints other than Agent Bricks Knowledge Assistant endpoints; custom retrievers, custom inputs and outputs, or fine-grained streaming control; custom Python logic between tool calls such as conditional branching or state management; and control over inference parameters such as temperature.

Read that list in reverse and it is a feature comparison.

Capability	Supervisor API (Beta, EOL 30 Sep 2026)	Custom agent loop on Databricks Apps
Agent loop	Managed by Databricks	You write it
Client-side function tools	Cannot mix with hosted tools	Supported
Inference parameters such as temperature	Not controllable	Controllable
Custom logic between tool calls	Not supported	Supported
Usage tracking	Not supported in the Beta	Standard Apps billing per compute hour

A team that hit any of those five limits was already planning to leave. A team that hit none of them has a genuine rewrite in front of it. The real cost here is usually the loop's error handling and the trace plumbing, not the happy path.

India-specific considerations

Indian teams running Databricks on Premium workspaces should treat the default-on Beta setting as an inventory problem before it becomes a migration problem. Where agent traffic touches personal data, the move from a managed endpoint to a self-written loop changes who is responsible for logging and retention of prompts and tool outputs, which is a Digital Personal Data Protection Act 2023 question about purpose limitation and storage, not only an engineering one. The Supervisor API stored OpenTelemetry traces in Unity Catalog by design; a hand-written loop stores whatever you tell it to store. Decide that deliberately rather than inheriting it.

What is still unknown

Databricks has not published a September 2026 release-notes page as of 23 August 2026, so there is no confirmation yet of whether the deprecation will appear there retrospectively. The notice does not say what happens to in-flight requests on 30 September, whether the endpoint returns an error or is removed outright, or whether workspaces with the preview enabled receive an in-product warning. The release-types page says a UI notification appears when GA is approaching for a Beta feature; it says nothing about a notification when a Beta feature is retired.

FAQ

When does the Databricks Supervisor API stop working?

Databricks has set end of life at 30 September 2026. Both the Supervisor API reference page and the tutorial page for building an app on it carry that date, and both were updated on 21 August 2026. After that date the documentation states the API will no longer be available.

Why is the deprecation not in the August 2026 release notes?

Databricks has not explained the omission. Checked on 23 August 2026, neither the AWS documentation nor the Azure Databricks mirror of the August 2026 platform release notes contains the word Supervisor, despite listing 66 other entries for the month, including many Beta features.

Is Agent Bricks Supervisor Agent also being deprecated?

No. Supervisor Agent is a separate Agent Bricks product that coordinates Genie Agents, agent endpoints, Unity Catalog functions and MCP servers, and it carries no deprecation notice. Only the Supervisor API, the managed agent loop reached at the OpenResponses-compatible endpoint, has an end-of-life date.

Could my workspace have the Supervisor API enabled without my knowing?

Yes, if you are on Premium. Databricks' release-types table states Beta features are on by default for Premium customers and off by default for Enterprise customers. Workspace and account admins control previews from the Previews page, so a proof of concept could have used the feature without a deliberate opt-in.

What is the recommended replacement?

The two documentation pages differ. The reference page recommends migrating to custom agents on Databricks Apps. The tutorial page is more specific and says to use a custom agent loop instead. The practical reading is that you keep Databricks Apps as the host and write the agent loop yourself.

Does the migration give me anything back?

Yes. Databricks documents five things the Supervisor API could not do: mix client-side function tools with hosted tools, use agent endpoints beyond Knowledge Assistant, support custom retrievers or fine-grained streaming, run custom Python logic between tool calls, and control inference parameters such as temperature. A custom loop supports all five.

How do I find affected code quickly?

Search for calls to the OpenResponses-compatible endpoint at ai-gateway/mlflow/v1/responses. Then look for a Databricks App whose agent.py makes a single API call rather than running a loop, which is the shape produced by the Supervisor API tutorial. Finally check the Previews page at both account and workspace scope.

Should Beta features be in production at all?

Databricks' own release-types table answers no: Beta carries no production recommendation, no interface stability, no SLA, and support routed to the engineering team rather than the support team. The gap between that policy and the on-by-default setting for Premium customers is how Beta code reaches production anyway.

How eCorpIT can help

Our senior engineering teams run platform-deprecation inventories of this kind: finding the calls, sizing the rewrite, and rebuilding a managed agent loop as maintained code with its own tracing and error handling. We work across Databricks, AWS Bedrock and the Claude and OpenAI APIs, so the replacement loop is not written against a single vendor's assumptions. eCorpIT is CMMI Level 5 and ISO 27001:2022 certified, and we design applications aligned with DPDP requirements for prompt and trace retention. Book a Databricks agent migration review and we will map your endpoint usage against the 30 September date.