OpenAI's Agent Builder is scheduled to shut down on November 30, 2026. If a production workflow still depends on it, I would avoid a direct screenshot-for-screenshot rebuild. The safer migration target is the workflow's behavior: inputs, decisions, tool calls, state, approvals, retries, and outputs.
The date and migration notice are in OpenAI's official Agent Builder documentation.
Here is a Node-RED migration checklist that keeps those behaviors explicit.
1. Convert the canvas into contracts
For every Agent Builder node, write down:
- accepted input schema
- output schema
- side effects
- retry policy
- timeout
- whether a human approval is required
This catches hidden assumptions before the old builder disappears.
2. Map responsibilities to subflows
A practical Node-RED shape is:
HTTP In -> Normalize -> Router -> Tool subflow -> Verify -> Approval gate -> HTTP Response
Keep each tool in its own subflow. The router should choose a capability, not embed credentials or business logic.
3. Add state that survives restarts
Store at least:
correlationIdidempotencyKey- current step
- tool receipt or external reference
- retry count
- approval status
- last error class
A larger context window is not a substitute for resumable state.
4. Separate failure classes
Use separate branches for:
- transport or HTTP failure
- invalid JSON/schema
- model refusal
- tool failure
- review candidate
Do not send all five into the same generic retry loop. Retries should have a budget, and exhausted items should go to a dead-letter flow.
5. Move secrets out of the flow export
Use Node-RED credentials or environment variables. A portable flow should export topology and configuration names, not API keys.
6. Test with fixtures before cutover
Create a small set of saved inputs for:
- normal request
- duplicate request
- delayed tool
- malformed model output
- approval denied
- restart during a tool call
Compare semantic outcomes, not wording.
7. Make model choice replaceable
Keep the provider adapter behind one subflow so the workflow does not change when a model changes. A unified API layer such as Tokatlas can be used to compare candidate models without rewriting the rest of the flow.
The useful migration artifact is not a copied canvas. It is a documented, testable set of contracts that can run after any one vendor UI is gone.
For people already migrating: which part is proving hardest to reproduce in Node-RED—state, tool calling, approvals, or evaluation?