Workbench and OpenAI Agent Builder
Workbench or OpenAI Agent Builder?
OpenAI shuts down Agent Builder and its Evals product on November 30, 2026, and points teams to the Agents SDK or to workspace agents in ChatGPT. Workbench gives your team one place to design agents on any model, test them with the people who know the work, and ship versions your product can count on.
Claim triage
Agent Builder · Version 3 · Shuts down Nov 30, 2026
Start, then Agent, then If / else, then End.
Agent: classify the claim into one of four kinds.
To keep it, rebuild it in the Agents SDK or as a ChatGPT workspace agent.
Claim triage
- Classifier with four kindsLive
- Tried on 40 recent claims92%
- Claims lead reacted by link3 changes kept
- ModelAny provider
The tests and the versions stay with the flow
What changes when you move.
- 1
Pick any model, and change it later
Choose from models across several providers, OpenAI's among them, or add your own provider. Switching a model is an edit, and your tests tell you whether it helped.
- 2
Tests that stay with the work
Caliper keeps your datasets, rubrics and reviewers, tests each version, and shows the score when you publish. OpenAI is retiring its Evals product on the same date as Agent Builder.
- 3
The people who know the work shape it
Send the claims lead a link to the draft, and their thumbs and comments become edits. Each edit reruns your tests and is undone if the score drops.
- 4
One place for builders and the rest of the team
The Agents SDK lives in your code, and workspace agents live in ChatGPT. Workbench keeps the visual editor, the published versions your code calls, and the share links your team uses in one place.
- 5
Processes that take days
Write a Task as a sentence, and Workbench turns it into steps, asks named people to approve or edit, escalates when they don't answer, and records the outcome in Ledger.
When Workbench is the right choice
- You want to keep a visual editor after Agent Builder shuts down.
- You want to choose models from more than one provider.
- You need tests and expert review on every version before it ships.
- The process runs for days and needs named people to sign off.
Job by job
Where it's built
OpenAI Agent Builder: Agent Builder until November 30, 2026, then the Agents SDK or ChatGPT.
Workbench: An editor for each kind of job, plus a canvas for anything custom.
Models
OpenAI Agent Builder: OpenAI's models.
Workbench: Models across several providers, or your own provider.
Testing
OpenAI Agent Builder: Trace graders and Evals, both retiring with Agent Builder.
Workbench: Caliper tests every version, and the score shows when you publish.
Versions
OpenAI Agent Builder: Versioned workflows your code calls by ID.
Workbench: Numbered versions your product pins, with the draft kept apart.
Where people use it
OpenAI Agent Builder: ChatKit in your app, or ChatGPT and Slack for workspace agents.
Workbench: A chat page by link, Slack, text, a webhook, or a Napkin button.
Running it yourself
OpenAI Agent Builder: Export the workflow as Agents SDK code.
Workbench: Export a flow and run it in your own Node process with the open-source SDK.
Moving off Agent Builder? Bring the workflow and its test cases.
Rebuild the agent in Workbench, keep OpenAI's models if they work for you, and test the new version against the cases you already trust.
- Paste the agent's instructions into a Workbench agent and pick the same OpenAI model.
- Bring your test cases into a Caliper dataset from a CSV file, and run them against each version.
- Keep ChatKit or ChatGPT for the chat, and call the published flow from your code.
Questions
- Can we keep using OpenAI's models?
- Yes. OpenAI's models are in the list, alongside other providers.
- What if we want to leave Workbench later?
- Export a flow and run it in your own Node process with the open-source SDK.
Try Workbench with your domain.
Bring the Agent Builder workflow you need to keep, and start there.
No website? Describe what you do
If you'd like help with the first changes, talk to us.