Updated November 12, 2026: OpenAI Agent Builder is deprecated and scheduled to shut down on November 30, 2026. Our earlier guide recommended the visual builder for new projects without explaining its retirement. This correction replaces that advice while preserving the original article URL and publication date.
Agent Builder was OpenAI’s visual canvas for composing agent workflows, a different product from the code-first Agents SDK. The practical question now is how to preserve useful workflow behavior without depending on the retiring builder.
What is shutting down, and when?
According to OpenAI’s deprecation schedule, Agent Builder shuts down November 30. Existing Evals became read-only October 31; the Evals dashboard and API are scheduled to close November 30. Reusable prompt objects and the v1/prompts API have the same November 30 deadline. ChatKit remains available. These are separate dependencies, so migrating a workflow alone does not settle every affected integration.
What should replace Agent Builder?
OpenAI’s migration guide describes exporting the workflow as Agents SDK code, then continuing with the SDK or recreating the workflow as a ChatGPT Workspace Agent. Export is not an automatic conversion into a Workspace Agent or a guarantee of equivalent behavior. Workspace Agents require an eligible Business, Enterprise or Edu workspace with the feature and creation permissions enabled.
Our recommendation: start from the application boundary. A customer-facing application that needs its own authentication, storage and release process should evaluate a server-owned runtime. An internal workflow that belongs in ChatGPT may fit Workspace Agents, subject to access and capability checks. Neither choice removes the need to test approvals and tool access.
Export before rebuilding
In the builder’s Code view, choose the Agents SDK export and the appropriate TypeScript or Python version. Keep the complete export. Treat it as migration material rather than a production-ready replacement: configure dependencies, credentials and the destination runtime, then test representative inputs. Recreating the workflow in Workspace Agents also means reviewing its instructions, triggers, tools, permissions and control flow.
Does ChatKit still work?
Yes. The current ChatKit guide describes a custom-server integration using the ChatKit Python SDK with an agentic backend, including the Agents SDK. The existing Agent Builder-hosted integration is a transition path, not the recommended foundation for new work. Your server must authenticate users and supply unique user identifiers.
Keeping the chat interface is not the same as keeping the hosted workflow. Inventory session creation, attachment handling, tool execution and conversation ownership separately. Test that one user cannot retrieve another user’s session after the backend changes.
A practical migration checklist
- Inventory live dependencies. Identify the workflows actually serving users, their callers, owners and external integrations. Include evaluation tooling and reusable prompt references.
- Preserve a baseline. Save the workflow export, configuration and permitted evaluation material. Capture representative successful and failed cases, excluding secrets and unnecessary personal data.
- Choose a destination. Write down which system will own state, tool authorization, background execution and incident recovery.
- Rebuild and test behavior. Include normal requests, ambiguous instructions, tool failures, approval rejection and malicious text inside retrieved documents.
- Validate the user boundary. Check authentication, conversation isolation, upload permissions and audit trails, not only answer quality.
- Plan the cutover. Define acceptance criteria, an owner, a deployment window and a fallback that does not depend on a service after its shutdown.
What about Evals and reusable prompts?
Do not confuse the retirement of OpenAI’s Evals product with the end of evaluation as an engineering practice. Preserve useful test cases and replace the affected execution path. OpenAI links to a Promptfoo migration route. For reusable prompts, its recommended direction is to move prompt content into application code. Review version references rather than assuming a copied prompt string preserves the surrounding behavior.
For your replacement test suite, measure task completion, unauthorized tool attempts, approval handling, latency and cost on the same representative inputs. Keep failures in the results. A migration that returns fluent text but loses an approval step is not ready.
How to avoid a rushed migration
Nerova’s recommendation is to migrate one representative workflow first, compare its behavior against the baseline, then expand. Record what changed and why. Prioritize production dependencies over unused prototypes, but do not delete historical configuration merely to simplify the checklist.
Do not start a new Agent Builder deployment because an older tutorial still describes it as the fast path. Check the current notice, select a supported destination and retain the workflow’s safety and operational requirements. For the code-first option, see our Agents SDK guide.