Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

OpenAI shuts Agent Builder, Evals, and prompts November 30

Written November 4, 2026

Three OpenAI surfaces close November 30, 2026. Where each workload moves — Agents SDK, Workspace Agents, Promptfoo, or your own code — and the export catch.

What shuts down, and when

On 3 June 2026, OpenAI notified developers that three products were being deprecated. All three are scheduled to shut down on 30 November 2026:

At shutdown, each surface stops being accessible. The migration for each one is different, and none of them is automatic — plan per surface, not as one move.

Agent Builder: export the workflow as code, then rebuild it

OpenAI's migration guide exports an existing workflow as Agents SDK code. Open the workflow in Agent Builder, select Code, select Agents SDK, pick TypeScript or Python, and copy the complete export. From there, two destinations are named:

The catch is stated in the guide itself: the export process does not convert your workflow graph and does not guarantee that every behaviour transfers unchanged. Review control flow, triggers, tools, and permissions as you test the migrated agent, and validate the application's configuration and behaviour before deploying it. Workflows with strong determinism at their core may not migrate faithfully to a workspace agent, and connected apps, authentication, publishing, and permission configuration need separate review in ChatGPT. A workflow that has not been tested end to end in its new home has not been migrated — it has only been copied.

Evals: recreate each eval by hand in Promptfoo

OpenAI recommends Promptfoo for continuing evaluation workflows, and its guide covers manually recreating an evaluation — it does not rely on an export feature. Recreate the evaluation's prompts, provider, test cases, and assertions in a promptfooconfig.yaml file, mapping graders configured as testing criteria in the eval to grading behaviour configured as assertions and metrics in Promptfoo. Then run a fresh evaluation and open the results viewer:

The new run is separate from any previously completed OpenAI Evals runs: past results do not carry over. After the first run, compare the recreated prompts, assertions, and outputs before using the eval in an ongoing workflow. Similarity-based scoring may not produce identical numerical results across systems, and any manually recreated grader — especially an LLM-as-a-judge grader — should be validated before you rely on it for regression decisions. A grader that has not been re-validated in Promptfoo is a new grader, not a moved one.

Reusable prompts: move the content into your application code

For reusable prompts, OpenAI's instruction is short: move reusable prompt content into your application code. There is no replacement prompt-object service named on the deprecations page. Copy each prompt's content out of the dashboard or API while it is still accessible, store it with the code that uses it, and change the code that read the prompt object to read the stored copy instead. Check every call site that referenced a prompt object by ID — a prompt stored in code but still read from v1/prompts will fail at shutdown exactly like one that was never moved.

Do not confuse this with the model removals

Two further dates on the same deprecations page are model removals, not product shutdowns, and they need separate migration work:

Moving a model and moving a workflow are different jobs. Finishing one does not finish the other.

Order of work

  1. Inventory all three surfaces now. List every Agent Builder workflow, every eval, and every reusable prompt object in use, including ones owned by former team members or embedded in scheduled jobs.
  2. Move Agent Builder workflows first. They need the most testing: export, rebuild, and run representative inputs in the new home before the canvas disappears.
  3. Recreate evals next, and re-validate graders. An eval moved without its graders re-validated will pass or fail for reasons nobody has checked.
  4. Move prompt content into code last, and grep for prompt IDs. The change is small per prompt, but a single missed call site fails at shutdown.
  5. Cut over before 30 November. The Evals read-only date has already passed; nothing in an existing eval can be edited now, only recreated elsewhere.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Move off Claude Sonnet 4.5 before it retires · GPT-6.1 Sol, Ultrafast, and the new Pro tiers