At DevDay in October 2025, OpenAI put a visual agent canvas on the big screen. Agent Builder was AgentKit's headline: drag a node for a tool, wire it to another for a handoff, add a guardrail, ship an agent without writing the orchestration yourself. It was the company's answer to a year of developers gluing together their own loops. Eight months later, on June 3, 2026, OpenAI added it to the deprecation tracker. Agent Builder shuts down November 30, 2026.

It does not go alone. The same notice retires the Evals platform — the managed place to run datasets against graders in the dashboard — and Reusable Prompts, the v1/prompts store that let you version a prompt and call it by ID. Evals goes read-only on October 31, so the practical deadline for anyone with evaluation history is a month earlier than the headline date: export before then or lose the ability to.

The migration is a retreat up the stack#

The interesting thing about a deprecation is where the vendor sends you next, because that's the product they actually believe in. For Agent Builder, OpenAI points to two places. One is the Agents SDK — code, not canvas, the same primitives you'd reach for comparing it against Pydantic AI or Google's ADK. The other is Workspace Agents inside ChatGPT, no-code but firmly on the consumer surface, not the developer one. Neither is a like-for-like replacement for a visual builder aimed at developers. A flow you assembled in the canvas doesn't export to SDK source; you re-write it.

The product they send you to is the product they actually believe in — and it's the one you were always able to build without them.

For Evals, the pointer is sharper. OpenAI's guidance references Promptfoo — an open-source evaluation tool the company does not own or operate. That is an unusual thing for a platform to do. When a vendor retires a feature, it normally folds the workflow into an adjacent first-party product. Handing eval runs to an outside project is closer to an admission: managed evals were not a business OpenAI wanted to keep staffing, and the observability-and-eval layer has enough good third-party options that OpenAI would rather cede it than maintain it.

What survives tells the story#

The cleanest read of any wind-down is the list of what's kept. Here it's short and coherent: the Responses API (/v1/responses) and ChatKit, the embeddable chat UI, are both untouched. OpenAI is holding the two ends — the raw API at the bottom and a drop-in chat widget at the top — and letting go of the middle: the visual builder, the eval dashboard, the hosted prompt store. Everything being retired was an attempt to own a layer above the API. Everything being kept is the API itself, or a thin veneer on it.

That middle layer is exactly where a lot of 2025's agent tooling lived, first-party and third-party alike. The bet was that developers wanted a managed place to compose agents, store prompts, and grade outputs without running infrastructure. What the retreat suggests is that the developers who are serious enough to pay ended up back in code — writing agents in an SDK, keeping prompts in their own repo, running evals in a tool they control — and the ones who wanted no-code were better served inside ChatGPT than in a separate developer console. The visual builder fell into the gap between those two audiences.

What to actually do before October 31#

If you have anything in these products, the order of operations is boring and time-boxed:

None of this is an emergency — you have until late November, and the API you're building on isn't going anywhere. But the direction is worth naming plainly. The no-code agent-builder era at OpenAI lasted about eight months, and it ended with the company telling its developers what many of them had already concluded: if the agent matters, you write it down as code.