By Greg Nowak. Updated 14 September 2026.
OpenAI is retiring Agent Builder on 30 November 2026. If it powers a customer conversation, an internal process or a client deliverable, put its migration on the delivery plan now. The work includes preserving the decisions, connections and controls that make the workflow useful.
A prompt might explain how to assess a request. It will not necessarily preserve who approves an action, what happens when a tool fails, or how an unfinished conversation resumes. Those details need to survive the move too.
What is retiring, and when?
OpenAI’s deprecation schedule confirms that Agent Builder shuts down on 30 November 2026. The Evals dashboard and API share that shutdown date, with existing evals becoming read-only on 31 October. ChatKit remains available.
That distinction matters if your website uses ChatKit: keeping the chat interface does not resolve its dependency on an Agent Builder workflow. Identify what runs behind the interface before deciding how much needs replacing.
Also check for reusable prompt objects. OpenAI lists their retirement, including the v1/prompts API, for 30 November. A replacement that still depends on those objects would leave another deadline unresolved.
Start with the live workflow and its owner
List each workflow, its business owner, published version, users and connected systems. Prioritise by what would stop working if it disappeared. An abandoned demonstration and an agent handling incoming enquiries deserve different budgets.
For agencies, record which client owns the account and credentials, who can approve a release, and who will maintain the replacement. Resolve access before development starts.
Next, take the export. OpenAI’s migration guide gives this route:
- Open the workflow in Agent Builder.
- Select Code in the top navigation.
- Choose Agents SDK.
- Select TypeScript or Python and copy the complete export.
Save it in version control with the workflow version and export date. Treat it as a starting point: OpenAI explicitly warns that migration does not guarantee unchanged behaviour.
| Preserve | Capture alongside the export | Acceptance check |
|---|---|---|
| Instructions | Prompts, variables, model settings and output formats | Representative requests produce usable results |
| Routing | Branches, handoffs, conditions and stopping rules | Each important path reaches the correct outcome |
| Tools | Input schemas, authentication dependencies, timeouts and retries | Valid, invalid and failed calls behave correctly |
| Controls | Approvals, guardrails and escalation recipients | Restricted actions wait for the right permission |
| State | History, session boundaries and pending work | Resumed conversations retain the necessary context |
| Evaluations | Test cases, scoring rules, known failures and baseline results | The replacement meets agreed acceptance criteria |
Choose a destination your team can operate
OpenAI points to the Agents SDK for workflows maintained through code and ChatGPT Workspace Agents for workflows built through natural language and shared with teams. Workspace migration requires an eligible Business, Enterprise or Edu workspace with agent access and creation permission.
My practical recommendation is to choose based on ownership and required behaviour:
- Choose the SDK when developers need explicit control over application integration, branching, testing and releases. Include hosting, monitoring and maintenance in the scope.
- Consider Workspace Agents when employees will maintain and supervise the process in ChatGPT. Check that the required tools, permissions and triggers can be recreated. Workflows requiring strictly predictable routing need particular care.
- Retire unused workflows when there is no continuing business need. Archive the useful material and confirm that nothing still calls them.
Before committing, rebuild one representative path, including a tool call and an approval or failure case. This exposes more migration effort than a successful text-only demonstration.
Make tool behaviour explicit
For custom function tools, OpenAI’s function-calling documentation separates the model’s request from execution: your application runs the function and returns its result. A tool description alone cannot preserve that application behaviour.
Validate arguments, enforce permissions where actions execute, set timeouts and limit retries. For actions that change records or send messages, prevent repeated requests from producing duplicate effects. Define what users see when a dependency is unavailable and who receives the escalation.
Assign ownership for secrets, deployment, logs and alerts. Include a short operating guide explaining how to investigate a failed run and pause the workflow.
Move the tests before changing the entry point
I would set an internal target before 31 October for preserving and recreating evaluation assets. That leaves time to check them before the read-only stage; it is a planning recommendation, not an earlier shutdown date.
OpenAI’s Promptfoo migration guide describes manually recreating prompts, test cases, providers and assertions in promptfooconfig.yaml. It does not depend on an Evals export feature. After installation and configuration, its example runs:
promptfoo eval -c promptfooconfig.yaml --no-cache
promptfoo viewCheck recreated graders before trusting their scores: results may differ between systems. Preserve historical results separately so the team can understand the original baseline.
Test normal requests, ambiguous inputs, tool failures, approvals and resumed sessions. Compare routes, tool arguments and business outcomes as well as answer quality. During replay or shadow testing, disable real writes or use test systems so comparison runs cannot repeat customer actions.
Switch with time to observe the replacement
Release to a limited group first. Agree acceptable quality, response time, cost and failure levels with the process owner. Expand only after reviewing actual runs.
Keep a fallback that remains usable after 30 November, such as a manual queue or reduced service. Agent Builder cannot be your rollback destination after shutdown. Schedule the main switch early enough to investigate problems while the original remains accessible.
If your team needs help scoping this work, I can help inventory the workflows, identify dependencies and plan the rebuild, testing and handover. Get in touch with a short description of what your agent does, which systems it touches and who relies on it. That gives us a useful starting point for estimating the work.
Related on GrN.dk
- A Voice Agent Is Only Ready When the Human Handoff Works
- AI Agents Need a Spending Brake, Not Just a Billing Dashboard
- AI automations need a spend dashboard before the first runaway bill
Need help with this kind of work?
Discuss your Agent Builder migration with Greg Get in touch with Greg.
