OpenAI's AgentKit and the eight-month lifespan of Agent Builder

What AgentKit actually bundled

The October 6, 2025 launch grouped several distinct tools under one name. Agent Builder was a visual canvas for composing multi-agent workflows with drag-and-drop nodes, versioning, guardrails, and inline eval configuration. ChatKit handled the frontend problem: embedding chat-based agent UIs that stream responses, manage threads, and show model reasoning. Connector Registry gave admins a single panel to govern data sources like Dropbox, Google Drive, SharePoint, and Microsoft Teams across ChatGPT and the API.

OpenAI also expanded Evals with four features — datasets, trace grading, automated prompt optimization, and third-party model support — and added custom tool calls and custom graders to the reinforcement fine-tuning beta, which was generally available on o4-mini and in private beta for GPT-5. The framing was consolidation: replace the fragmented mix of orchestration, connectors, manual eval pipelines, and frontend work with one integrated stack built on the Responses API.

The update that reframes the whole page

At the top of the same announcement sits a note dated June 3, 2026: OpenAI is winding down Agent Builder and Evals, and both will be gone from the platform after November 30, 2026 — roughly eight months after they shipped. That is the single most important fact on the page, and it appears above the launch copy that still describes both products in the present tense.

OpenAI names two replacement paths. Workflows meant to run as code should move to the Agents SDK. Use cases suited to natural-language prompting should move to Workspace Agents in ChatGPT. In effect, the visual-canvas approach and the standalone eval product are being split between a code path and a ChatGPT-native path, with the middle-layer builder tools retired.

Update on June 3, 2026: OpenAI is winding down the Agent Builder and Evals products. From November 30, 2026 onward, they will no longer be available on the OpenAI platform.Montana Labs

The customer numbers that carried the launch

The original announcement leaned heavily on named-customer metrics. Ramp said Agent Builder cut iteration cycles by 70% and shipped an agent in two sprints instead of two quarters. LY Corporation built and ran a multi-agent workflow in under two hours. Canva said ChatKit saved over two weeks of frontend work with integration in under an hour. Carlyle credited the Evals platform with cutting development time on a due-diligence framework by over 50% and raising accuracy 30%. Broader references cited Klarna handling two-thirds of tickets and Clay's 10x growth via a sales agent.

The tension is that several of these testimonials specifically praise Agent Builder — the tool now being deprecated. Ramp and LY Corporation both attributed their speed to the visual canvas keeping engineers, subject-matter experts, and other stakeholders in one interface. Whatever those teams built through the canvas will need to migrate to the Agents SDK or ChatGPT Workspace Agents before the November 2026 cutoff.

What the wind-down tells builders about betting on early platform tools

The concrete implication is migration risk on OpenAI's own first-party tooling. Agent Builder promised to compress months of orchestration into hours, and its adopters got that speed — but the abstraction itself proved short-lived, while the underlying primitives it sat on (the Responses API and Agents SDK, released in March 2025) are the recommended survivors.

For teams choosing where to invest, this launch-then-retract cycle is a useful data point: the durable layer was the code-level SDK, not the visual builder layered on top of it. ChatKit and the new Evals capabilities were made generally available while Agent Builder and Connector Registry shipped only in beta — and it was the beta-stage visual builder, plus Evals, that got cut. Anything a team standardized on Agent Builder now needs a second build, this time against the SDK OpenAI is pointing everyone toward.