OpenAI Agent Builder and Evals Shut Down 30 November: Your Move

OpenAI's Agent Builder and Evals close on 30 November 2026, forcing a code-first rethink of how you build and test agents.

If you built an agent workflow on OpenAI's drag-and-drop Agent Builder, or you lean on its hosted Evals to grade model outputs, the clock is now down to weeks. Both products go away on 30 November 2026, and one of the two evaluation surfaces has already gone read-only. This is not a rumour or a roadmap slide — it is on OpenAI's own deprecations page, and it is the first time a frontier lab has torn down a shipped low-code agent product this completely. Here is what's changing, why it matters even if you never touched Agent Builder, and how the main alternative frameworks stack up if you're rebuilding anyway.

What's actually shutting down, and when

On 3 June 2026, OpenAI added three platform-level deprecations to its official deprecation tracker, all with a common shutdown date of 30 November 2026: the Evals Platform, the Agent Builder, and the Reusable Prompts (v1/prompts API). The Evals side moved in two stages: existing evals became read-only on 31 October 2026, and the Evals dashboard and API are scheduled to shut down on 30 November 2026. Agent Builder follows the same final date: deprecation for Agent Builder was announced on 3 June 2026, and Agent Builder is scheduled to shut down on 30 November 2026.

One useful carve-out: ChatKit, the embeddable chat interface that shipped alongside Agent Builder, is not part of this wind-down. ChatKit is the one AgentKit component that is not affected by the Agent Builder deprecation, and OpenAI's documentation confirms it remains available going forward. The open-source evals repository on GitHub is also untouched — this deprecation is specifically about the hosted dashboard and API.

Why this matters even if you skipped Agent Builder

Agent Builder was never a standalone product. OpenAI AgentKit bundled a visual Agent Builder, ChatKit embeddable chat UI, Connector Registry, Guardrails, and Evals for building agentic workflows on OpenAI's models. It launched alongside the rest of AgentKit at DevDay in October 2025, pitched as the low-code front door into OpenAI's agent stack. Retiring the visual builder and the hosted evaluation dashboard after barely a year on the market is a strong signal about where OpenAI wants developers to live: in code, not in a canvas.

OpenAI has said as much directly. For workflows that should continue as code, OpenAI recommends the Agents SDK, and for use cases better suited to natural language prompting, it recommends Workspace Agents in ChatGPT. If your team prototyped in Agent Builder because it looked like the sanctioned, lowest-effort path, that path is closing. The practical lesson for anyone evaluating agent tooling now is to weight long-term maintenance commitment, not just day-one polish — a lesson worth remembering before standardising on any other vendor's visual builder.

Migration paths for each piece

Agent Builder workflows. Export what you can and move to the Agents SDK for anything that needs to run unattended in production, or to ChatGPT Workspace Agents if the workflow is closer to a guided, natural-language task. OpenAI's own migration guide points to exactly this split.

Evals. OpenAI's deprecation notice links to its cookbook guide, "Moving from OpenAI Evals to Promptfoo", as the migration path. Promptfoo is open source and framework-agnostic, so grading logic you build there survives the next vendor pivot too. If your evals already cover agent and tool-use scenarios rather than single prompt-response checks, a repo-owned eval harness tends to be a better long-term fit than any hosted dashboard.

Reusable prompts. The v1/prompts API and reusable prompt objects are scheduled to shut down on 30 November 2026; to migrate, move reusable prompt content into your application code. This is the simplest of the three changes — treat prompts as versioned code artefacts rather than server-side objects.

Checklist before 30 November

Inventory every Agent Builder workflow still in production and classify each as "rebuild in Agents SDK" or "rebuild in Workspace Agents."
Export Evals datasets and graders now; the dashboard has been read-only since 31 October, and waiting risks losing context on test intent.
Stand up Promptfoo (or an equivalent repo-owned eval setup) and port your highest-value regression checks before removing API calls to the Evals endpoint from CI.
Search your codebase for calls to the reusable prompts API and inline that content into your application.
Confirm every migrated agent runs end-to-end before the cutover, rather than treating 30 November as the day migration begins.

If you're rebuilding, how the main frameworks differ

Moving off Agent Builder is a natural point to ask whether the OpenAI-only Agents SDK is still the right fit, or whether a model-agnostic framework serves you better.

OpenAI Agents SDK — open source, code-first, and wired natively to OpenAI's models. It ships as openai-agents for Python (version 0.24.0, released 15 October 2026) and @openai/agents for TypeScript (version 0.20.0, released 8 October 2026). Best if you're committed to OpenAI models and want the shortest path from Agent Builder.
LangGraph — a graph-based orchestration runtime rather than a full application framework. Graph-based solutions like LangGraph give you precise control, suited to complex agents and production systems that need fault tolerance. It is model-agnostic and works well when you need durable, auditable multi-step execution.
CrewAI — best if you need role-based multi-agent prototypes up and running quickly with an intuitive mental model, and it is MIT-licensed with a large download base, making it a common choice for teams that want fast iteration over fine-grained control.
Microsoft Agent Framework (MAF) — MAF unifies Semantic Kernel and AutoGen into one .NET and Python SDK, with GA shipping on 3 April 2026, adding native MCP and agent-to-agent protocol support. It's the obvious pick for .NET or Azure-centred teams, and existing Semantic Kernel plugins carry across with modest rework.

None of these lock you into a single model vendor the way AgentKit's visual layer did, which is itself a reason to prefer a code-first framework over a hosted low-code builder going forward — the underlying logic is portable if the vendor changes course again.

Conclusion

The Agent Builder and Evals shutdown is a reminder that low-code agent tooling from any vendor can be withdrawn faster than the infrastructure you build on top of it. The safer long-term bet, confirmed by OpenAI's own guidance, is to keep agent logic and evaluation pipelines in code you own, using an open framework rather than a vendor dashboard. If you're comparing LangGraph, CrewAI, Microsoft Agent Framework, the OpenAI Agents SDK, or MCP servers for the migration, Intellibotz's directory is a reasonable place to start narrowing the options.