Direct answer: OpenAI has scheduled Agent Builder, the Evals dashboard/API, and reusable prompt objects including v1/prompts to shut down on November 30, 2026. Existing Evals became read-only on October 31. ChatKit remains available, but it does not remove the need to replace an Agent Builder-hosted backend. Start by exporting Agent Builder workflows, copying prompt content into your repository, and recreating critical evals in Promptfoo.[1]
15-minute risk check
Run this inventory before choosing a replacement:
# Reusable prompt IDs and prompt-object calls
rg -n 'pmpt_|v1/prompts|prompt\s*:\s*\{' .
# Agent Builder workflow IDs or ChatKit references
rg -n 'workflow[_-]?id|agent-builder|chatkit' .
# Hosted Evals IDs, API calls, datasets, or graders
rg -n 'evals?|grader|dataset' .
Also check deployment secrets, CI variables, no-code automations, serverless functions, and internal runbooks. A clean code search does not prove you are unaffected if a workflow ID lives only in a hosted configuration.
What to do for each match
- Agent Builder: export the complete workflow as Python or TypeScript Agents SDK code, then choose an application-hosted Agents SDK implementation or a ChatGPT Workspace Agent.
- Evals: preserve the test cases, prompts, graders, expected outputs, and useful historical results; manually rebuild the important suite in Promptfoo and compare behavior.
- Reusable prompts: copy prompt text into named, version-controlled builder functions and replace prompt IDs with explicit
inputandinstructionsin your application. - ChatKit: confirm what backend the interface calls. ChatKit remains available, but an interface pointing to an Agent Builder workflow still needs a replacement backend.[1]
Shutdown timeline
- June 3, 2026: OpenAI announced the three deprecations.
- October 31, 2026: existing Evals became read-only.
- November 30, 2026: Agent Builder, the Evals dashboard/API,
v1/prompts, and reusable prompt objects are scheduled to shut down.[1]
OpenAI’s notice gives the date but not a cutoff time or time zone. Treat November 30 as the outside limit, not a deployment target.
Path 1: migrate Agent Builder
OpenAI’s official guide says to open the workflow, select Code, choose Agents SDK, select TypeScript or Python, and copy the complete export. It also warns that the export does not convert the workflow graph or guarantee unchanged behavior.[2]
Choose the destination
Use the Agents SDK when:
- the workflow runs inside your product or backend;
- you need deterministic control over deployment, authentication, logging, retries, or approvals;
- the workflow is triggered by an API, webhook, queue, or schedule;
- you need to test and release it with application code.
Consider a ChatGPT Workspace Agent when:
- people start the workflow interactively;
- natural-language use and team sharing are the main requirements;
- the workspace is Business, Enterprise, or Edu and has permission to create agents.
OpenAI identifies those as the two migration paths. Workspace Agents can require manual recreation, and connected apps, authentication, publishing, permissions, tools, and safety controls need a separate review.[2]
Agent Builder export checklist
Before the old workflow becomes unavailable, save:
- the Python and/or TypeScript export;
- screenshots of the full graph and every node setting;
- instructions, model choices, structured-output schemas, and fallback paths;
- tool and MCP connection names, scopes, and non-secret configuration;
- trigger rules, input/output examples, error handling, and approval steps;
- a small set of representative successful and failed runs.
Do not put API keys or access tokens into the export archive. Record the secret names and required scopes, then restore values through your normal secret manager.
Validation gate
Test normal, edge, and denied-action cases against the replacement. Compare tool calls, outputs, permissions, latency expectations, and failure handling. OpenAI specifically recommends reviewing control flow, triggers, tools, permissions, configured capabilities, authentication, and representative inputs before creating or deploying the migrated agent.[2]
Path 2: move OpenAI Evals to Promptfoo
OpenAI recommends Promptfoo for continuing evaluation workflows. The official cookbook describes a manual recreation: map prompts, providers/models, test cases, graders, and scoring criteria into a portable configuration file, then run the suite locally or in CI.[3]
Preserve before rebuilding
Save the assets that would be hard to reconstruct:
- test inputs and reference answers;
- grader definitions and scoring thresholds;
- prompts and model/provider settings;
- expected tool calls or structured outputs;
- representative passing and failing outputs;
- historical results used for release decisions;
- notes explaining why each test exists.
A score without the underlying cases, grader, and model settings is weak migration evidence.
Minimal Promptfoo workflow
Create a promptfooconfig.yaml that includes the provider, prompt, tests, and assertions. Then validate and run it:
promptfoo validate config -c promptfooconfig.yaml
promptfoo eval -c promptfooconfig.yaml --no-cache
promptfoo view
Those are the validation and run patterns in OpenAI’s migration cookbook. The new run is separate from previous OpenAI Evals runs.[3]
Do not compare only the headline score
OpenAI warns that similarity-based scoring may not produce identical numerical results across systems. Manually recreated graders, especially LLM-as-a-judge graders, should be validated before being used for regression decisions.[3]
Use a small calibration set with known pass/fail expectations. Inspect disagreements, adjust the assertion or rubric only when justified, and keep the old and new evidence together. The goal is not to reproduce an arbitrary number; it is to preserve the release decision your eval was designed to support.
Path 3: move reusable prompt objects into code
OpenAI’s migration guide says to move managed prompt content into application code, replace prompt variables with function arguments, pass messages through input instead of referencing the prompt object, and use repository history, pull-request review, tests, and evals for versioning.[4]
Recommended repository structure
src/
prompts/
supportReply.ts
leadQualification.ts
summarizeDocument.ts
evals/
supportReply.yaml
fixtures/
Each prompt builder should have:
- a stable function name;
- typed or validated variables;
- an explicit system/instruction section;
- a model setting owned by configuration;
- fixtures for the most important behavior;
- a rollback path through version control or a feature flag.
Avoid scattering large prompt strings across route handlers. OpenAI recommends a small prompts module with named builder functions and lightweight eval fixtures.[4]
Cutover without a blind switch
- Inventory every
pmpt_ID and the code path that calls it. - Copy the current prompt text and variables into the repository.
- Add tests for variable substitution and required instructions.
- Send a small test cohort through the code-owned prompt.
- Compare output quality, tool behavior, cost, and error handling.
- Deploy behind a feature flag if the workflow is business-critical.
- Remove the prompt-object dependency only after production verification.
One-page migration tracker
Copy this into an issue or spreadsheet:
Asset:
Owner:
Current dependency: Agent Builder / Evals / prompt object
Production call site:
Replacement:
Export or backup complete: yes/no
Secrets and permissions mapped: yes/no
Representative tests captured: yes/no
Replacement tested: yes/no
Rollback tested: yes/no
Cutover date:
Final verification evidence:
Prioritize assets by business impact, not by how easy they are to migrate. A low-volume workflow that changes customer records deserves more attention than a high-volume internal summarizer.
Common migration mistakes
Treating the Agent Builder export as a finished application
The export is a starting point. OpenAI says the graph is not converted and behavior is not guaranteed to transfer unchanged.[2]
Rebuilding evals without preserving the original evidence
Keep representative outputs and grader logic. Otherwise a passing new suite may measure a different behavior.
Moving prompts into code without tests
Version control records changes; it does not prove those changes are safe. Add fixtures and eval checks before the first cutover.
Assuming ChatKit means Agent Builder can stay
ChatKit remains available, but any ChatKit implementation backed by a retiring Agent Builder workflow still needs a new server-side or workspace-agent path.[1][2]
Waiting for November 30
The published date is a scheduled shutdown date, not a maintenance window. Finish the cutover early enough to observe real traffic and roll back.
FAQ
Is all of OpenAI’s agent stack shutting down?
No. The notice covers Agent Builder, the hosted Evals product, and reusable prompt objects. OpenAI says ChatKit remains available and directs Agent Builder users toward the Agents SDK or ChatGPT Workspace Agents.[1][2]
Can I still read existing OpenAI Evals?
OpenAI’s timeline says existing Evals became read-only on October 31, 2026, before the dashboard and API are scheduled to shut down on November 30.[1]
Does Promptfoo automatically import every OpenAI Eval?
The official cookbook describes manually recreating an evaluation and says that process does not require an OpenAI Evals export feature.[3]
Will an exported Agent Builder workflow behave exactly the same?
Not necessarily. OpenAI says the export does not convert the workflow graph or guarantee unchanged behavior, so manual review and representative testing are required.[2]
Where should reusable prompts live?
OpenAI’s migration path is to store prompt content in application code, make variables explicit, pass messages directly through the Responses API, and manage review and versioning with normal engineering controls.[4]
Sources
[1] https://developers.openai.com/api/docs/deprecations — Deprecations | OpenAI API
[2] https://developers.openai.com/api/docs/guides/agent-builder/migrate-from-agent-builder — Migrate from Agent Builder | OpenAI API
[3] https://developers.openai.com/cookbook/examples/evaluation/moving-from-openai-evals-to-promptfoo — Moving from OpenAI Evals to Promptfoo | OpenAI
[4] https://developers.openai.com/api/docs/guides/prompting/migrate-from-prompt-object — Migrate from prompt objects | OpenAI API