OpenAI's Agent Builder Shuts Down on 30 November. Here Are the Only Two Exits.

September 14, 2026. If you or a client built an agent workflow in OpenAI's Agent Builder, the drag and drop canvas, you have 77 days. OpenAI's deprecations page schedules Agent Builder to shut down on 30 November 2026, along with reusable prompt objects and the Evals dashboard and API. None of this is new news, and that is the problem: the notice went out on 3 June 2026, the deadline has been sitting quietly in a documentation table all summer, and a canvas that still opens and still runs today gives no warning that it stops in November.

What shuts down, and when

1. Agent Builder, 30 November 2026. OpenAI's wording: "On June 3, 2026, we notified developers using Agent Builder that the product is being deprecated. ChatKit remains available." The migration note directs you to "continue with the Agents SDK or ChatGPT Workspace Agents."
2. Reusable prompt objects and the v1/prompts API, 30 November 2026. Announced the same day, with a blunt migration path: "move reusable prompt content into your application code."
3. The Evals platform, in two steps. Existing evals become read only on 31 October 2026, and the Evals dashboard and API shut down on 30 November 2026. The 31 October date is the one that bites first, because a read only eval cannot be rerun to prove a migration did not change behaviour.

Note the order those dates fall in. If you intend to use your existing evals to verify that a rebuilt agent still behaves the way the old one did, you have until 31 October, not 30 November.

The exits OpenAI actually names

Two, and only two, for Agent Builder: the Agents SDK, which is code rather than canvas, or ChatGPT Workspace Agents, which keeps a no code surface but moves the agent inside the ChatGPT workspace. ChatKit, the embeddable chat frontend, is explicitly unaffected. It is worth being precise here because the two exits are not equivalent: one is a developer migration with a repository and a deployment, the other is a product change that alters where your client's users go to use the thing you built.

What it means for operators

The reason a 77 day deadline is worth a calendar entry rather than a shrug is that visual builders are where agency work accumulates fastest and gets documented least. A canvas built during a discovery call, handed over in a screen recording, and never committed to a repository has no changelog to warn anybody. The audit is short: list every client whose automation has an Agent Builder node in it, note whether the prompts live in the canvas or in code, and decide per client whether the exit is the Agents SDK or Workspace Agents. Agencies that resell automation under their own brand, the pattern we describe in white label SaaS for agencies, carry this risk twice over, because the client bought your product name and will bring the outage to you rather than to OpenAI.

There is also a planning lesson with a longer shelf life than this particular deadline. OpenAI publishes its notice periods: at least six months for generally available models, at least three months for specialised variants, and for preview products, retirement "with much shorter notice, such as 2 weeks." Agent Builder got roughly six months. A product still carrying a preview or beta label gets two weeks, which is not enough time to rebuild a client workflow. When we scope an AI automation build, the label on the surface it runs on is a line item, because it is the difference between a quarter of warning and a fortnight of it.

The cheap habit that would have made today a non event: keep prompts and agent logic in version control even when the vendor offers to keep them for you. The migration note for reusable prompts, "move reusable prompt content into your application code," is the whole lesson compressed into eleven words.