What is OpenAI Agent Builder? What it does and what replaces it
OpenAI Agent Builder is a visual canvas for agent workflows that OpenAI is shutting down on November 30, 2026. What it does and how to migrate to the Agents SDK.
OpenAI Agent Builder is a visual canvas for building multi-step agent workflows: you start from a template or a blank canvas, connect nodes, and define typed inputs and outputs (docs). It launched as part of OpenAI’s AgentKit (announcement). OpenAI says it is deprecating Agent Builder and that the product is scheduled to shut down on November 30, 2026 (docs). If you have a workflow in it, export it before that date.
What Agent Builder does
Each connection between nodes becomes a typed edge. You can run the workflow in preview with sample files and watch each node execute. Autosave runs continuously, and publishing creates a new major version that acts as a snapshot, which you can pin in API calls (docs). From the Evaluate tab you select traces and run custom trace graders on them.
The node reference lists the building blocks:
| Group | Nodes | What they do |
|---|---|---|
| Core | Start, Agent, Note | Start defines workflow inputs; Agent runs a model with instructions, tools and model configuration; Note holds comments and does nothing in the flow |
| Tools | File search, Guardrails, MCP | File search queries vector stores on the OpenAI platform; Guardrails check for PII, jailbreaks, hallucinations and other misuse and are pass/fail by default; MCP calls OpenAI connectors or third-party servers |
| Logic | If/else, While, Human approval | Branching and looping use Common Expression Language expressions; Human approval pauses for end-user review |
| Data | Transform, Set state | Transform reshapes data between nodes; Set state saves global variables |
These map onto ideas covered in what an agentic workflow is and what AI guardrails are. For the protocol behind the MCP node, see Model Context Protocol.
The shutdown timeline
| Date | What happens | Source |
|---|---|---|
| June 3, 2026 | OpenAI notifies developers that Agent Builder is deprecated | deprecations |
| October 31, 2026 | Existing evals on the OpenAI Evals platform become read-only | deprecations |
| November 30, 2026 | Agent Builder shuts down, as do the Evals dashboard and API, and the v1/prompts API and reusable prompt objects |
deprecations |
The deprecations page says “ChatKit remains available.” The Agent Builder docs say existing users can keep using it during the transition (docs).
How to move off Agent Builder
OpenAI documents two paths (migration guide). In both, you start by opening the workflow, selecting Code, choosing Agents SDK, picking TypeScript or Python and copying the full export. OpenAI states that this “does not convert your workflow graph or guarantee that every behavior transfers unchanged,” so plan to test.
| Path | What you do | Watch for |
|---|---|---|
| Agents SDK | Paste the export into your application, install the matching SDK and validate behavior in your runtime | You now own deployment, storage and approvals, which the SDK leaves to your application (agents guide) |
| ChatGPT Workspace Agents | Create a workspace agent in a ChatGPT Business, Enterprise or Edu workspace, paste the export with OpenAI’s prompt, review the generated instructions and test in Preview | Workflows that depend heavily on determinism “may not migrate faithfully to a workspace agent” |
If you used ChatKit as the chat front end, it stays. Its advanced integration runs on your own infrastructure and can “connect to any agentic service, including one built with the Agents SDK,” with the ChatKit Python SDK covering the backend (ChatKit). The Agents SDK itself is covered in what the OpenAI Agents SDK is. If you would rather write the loop yourself, see how to build AI agents.
What to carry over besides the graph
The shutdown dates also cover pieces outside the canvas:
- Evals: the Evals dashboard and API shut down on November 30, and existing evals are read-only from October 31 (deprecations). The migration guide does not name a replacement for evals or trace grading, so keep your test cases and graders in your own repo or an evaluation tool. See LLM evaluation.
- Prompts: reusable prompt objects and the
v1/promptsAPI shut down on the same day (deprecations), so keep prompt text in version control. - Safety settings: OpenAI’s guide says to pass untrusted input through user messages, use fixed schemas and enums between nodes, and enable tool approvals for MCP tools, including reads and writes (safety guide). The export does not promise to carry guardrail behavior over, so re-check these controls in the new runtime.
Tracing a migrated agent in HoneyHive
HoneyHive has no Agent Builder integration, and the HoneyHive docs do not mention Agent Builder or AgentKit. Once a workflow runs on the Agents SDK, it can be traced like any other SDK app: HoneyHive documents tracing agent runs, handoffs and tool calls with an OpenInference instrumentor (integration docs). That gives you traces to compare before and after migration. For what to capture, see what agent observability is.
FAQ
Is OpenAI Agent Builder being discontinued?
Yes. OpenAI notified developers on June 3, 2026 that it is deprecating Agent Builder, and the product is scheduled to shut down on November 30, 2026 (deprecations).
Can I export an Agent Builder workflow?
Yes. Open the workflow, select Code, choose Agents SDK, pick TypeScript or Python and copy the export. OpenAI says this does not convert your workflow graph or guarantee that every behavior transfers unchanged (migration guide).
Does ChatKit stop working when Agent Builder shuts down?
No. OpenAI says ChatKit remains available. Its advanced integration runs on your own infrastructure and can connect to any agentic service, including one built with the Agents SDK (ChatKit).
What are the alternatives to Agent Builder?
OpenAI's own migration paths are the Agents SDK and ChatGPT Workspace Agents (migration guide). Code-first frameworks are another route; see LangGraph and Pydantic AI.