A quick, unsentimental heads-up for anyone who built an AI workflow inside OpenAI’s Agent Builder. It is going away. On 3 June 2026 OpenAI announced it will wind down three things: the Evals dashboard, the visual Agent Builder canvas and the reusable Prompts API. Evals goes read-only on 31 October 2026, and Agent Builder, Evals and the prompts API all close on 30 November 2026. As things stand in late July, you have roughly four months. Do not leave it to November.
Why this keeps happening
Agent Builder was a drag-and-drop canvas for stitching together AI agents with guardrails and preview runs. It launched to fanfare. It is being switched off about a year later. This is the pattern with rented tools: the platform giveth a slick visual builder, the platform taketh away when it decides the feature is not worth maintaining. If your business logic lives inside someone else’s canvas, your business logic is on loan.
None of this makes OpenAI a villain. It happens everywhere. The lesson for a small operator is not ‘trust nobody’, it is ‘keep the important stuff somewhere you control’.
Where OpenAI wants you to go
The official off-ramp is the Agents SDK, OpenAI’s code-based way to build agents, which had a substantial update in April 2026 adding native sandbox execution and a model-native harness for longer-running work. There is an export path from Agent Builder to the SDK, so you are not rebuilding from a blank page. OpenAI also points workspace users toward its natural-language workflow product, but the durable answer here is code.
The move, in order
One, export everything now. Get your Agent Builder workflows out while the export tool still exists, not on 29 November.
Two, put your prompts and your test cases in version control you own. A plain Git repository will do. The reusable Prompts API is dying, so any prompt that only lives on OpenAI’s servers needs a copy on your side.
Three, decide where the logic actually lives long term. The Agents SDK is the low-friction choice if you are already all-in on OpenAI. If this whole episode has made you twitchy about lock-in, a self-hostable tool like n8n or a portable framework keeps your workflow yours, and lets you swap the model underneath without rebuilding the plumbing.
Four, rebuild your evals somewhere permanent. Losing the Evals dashboard is the real sting, because the tests that prove your agent works are the last thing you want to lose. Keep them as code.
The point is not panic. Four months is plenty. The point is that a deadline like this is a free reminder to own the parts of your stack that matter, so the next time a big company retires a button, it is an afternoon of admin and not a fire.
Did you know: OpenAI’s Evals dashboard goes read-only a full month before it closes, on 31 October 2026, so you can still read your old scores after you can no longer run new ones.