OpenAI is winding down Agent Builder and Evals — your migration guide

TL;DR — what changed and by when

• Two products, not all of AgentKit. On June 3, 2026 OpenAI deprecated Agent Builder (the visual workflow canvas) and the Evals platform. Per the deprecation notice, “ChatKit remains available” — the embeddable chat UI toolkit is not part of this wind-down.
• Final date: November 30, 2026. Both products leave the platform that day. Evals additionally goes read-only on October 31, 2026, so export your datasets before then.
• Where OpenAI points you. For Agent Builder: the code-first Agents SDK or Workspace Agents in ChatGPT. For Evals: the deprecation guidance references migrating to Promptfoo.
• The lesson. A hosted visual builder is a proprietary surface a vendor can sunset on its own schedule. Your tool integrations don’t have to ride along — if they speak MCP, they port to any runtime. More in the hedge section.

What AgentKit was, briefly

OpenAI announced AgentKit at DevDay on October 6, 2025 as, in Sam Altman’s framing, “a complete set of building blocks… designed to help you take agents from prototype to production.” It bundled four pieces:

• Agent Builder — a visual, drag-and-drop canvas for composing multi-agent workflows, built on top of the Responses API, with preview runs, inline eval configuration, guardrails, and versioning. This is the piece most people meant when they said “AgentKit.”
• ChatKit — an embeddable, brandable chat UI toolkit for putting an agent in your own product. Not deprecated.
• Connector Registry — an admin panel for managing how data and tools (pre-built connectors plus third-party MCP servers) connect across ChatGPT and the API. Not named in the deprecation notice.
• Evals for agents — trace grading, datasets for component-level assessment, automated prompt optimization, and the ability to run evaluations on external models. The hosted Evals product is the second thing being wound down.

The pitch was “prototype to production without writing the orchestration yourself.” Agent Builder was the headline: a Canva-like canvas for agent logic. Eight months on, that headline is the thing going away.

What’s being deprecated — precisely

Precision matters here, because “OpenAI killed AgentKit” is the wrong headline and it’ll send you migrating things that don’t need migrating. Per OpenAI’s deprecations page and the Agent Builder deprecation notice:

The exact timeline

Three dates to put on a calendar:

OpenAI says existing users can keep using both products through the transition window — nothing breaks the day the notice goes up. But the read-only cutover for Evals lands a month before the full shutdown, so treat October 31 as your real export deadline.

What to do now

If you have anything in Agent Builder or the Evals dashboard, here’s the order of operations.

If you use Agent Builder

1. Export everything. Pull your workflows, prompts, guardrail configs, and tool wiring out of the canvas while it’s still live. Treat the export as your spec.
2. Re-express the workflow as code. Each canvas node — an agent, a tool call, a handoff, a guardrail — maps to an object in the Agents SDK. It’s the same Responses API underneath, so the model and tool behavior should carry over; what you’re rebuilding is the orchestration the canvas was doing for you.
3. Re-wire connectors and MCP servers. The tools your agent called don’t change. If you connected third-party MCP servers, point the SDK at the same ones.
4. Keep your ChatKit frontend. ChatKit stays; swap its backend from OpenAI-hosted Agent Builder to your own Agents SDK service.

If you use the Evals platform

1. Export before October 31, 2026. After that, evals go read-only and you can’t edit or re-run them on the platform.
2. Pick a replacement harness. OpenAI’s guidance references Promptfoo; plenty of teams also run their own eval scripts. The important thing is owning your datasets and graders outside a single vendor’s dashboard.

Your options now

Strip away the branding and there are really two directions to go: stay closer to OpenAI’s code-first stack, or build on a framework-agnostic, MCP-based stack you can move between vendors. Both are legitimate. They optimize for different risks.

A nuance worth stating plainly: choosing the Agents SDK isn’t a wrong answer. It’s code, it’s open-source, and it speaks MCP, so it’s far less locked-in than the visual builder was. The distinction between A and B is really about how much of your orchestration is willing to be OpenAI-shaped versus vendor-neutral.

The hedge: why teams build on MCP

The durable takeaway isn’t “OpenAI deprecates things.” Every platform sunsets products; an eight-month cycle is just unusually visible. The takeaway is about which layer of your stack absorbs the blow when it happens.

A hosted visual builder is a proprietary surface. Its file format, its node semantics, its hosting are the vendor’s to change or retire. When Agent Builder leaves, the orchestration you expressed in it has to be re-expressed somewhere else — that work is unavoidable.

What doesn’t have to be redone is your tool layer, if it speaks an open standard. The Model Context Protocol is a vendor-neutral way for agents to talk to tools and data. A tool exposed over MCP works the same whether it’s called from the Agents SDK, from LangGraph, from Claude, or from a client that doesn’t exist yet. That’s the hedge: keep the orchestration framework swappable and the tool integrations on an open protocol, so a deprecation costs you a framework migration, not a from-scratch rebuild of every connector.

This is also why the Connector Registry and MCP servers came through this episode untouched. Standards outlive the products built on top of them. If you’re re-architecting anyway, it’s the right moment to push tool access down onto MCP and treat the agent framework — OpenAI’s or anyone else’s — as the replaceable part. For securing those connections across machines, see our guide to secure MCP tunnels; for how the consumer-facing agent products stack up, see Cowork vs ChatGPT Agents vs Operator vs Manus.

Community signal

The deprecation was widely reported by practitioners within hours. The most common framing pointed straight at the migration options rather than the drama:

That’s the right altitude to take it from. The news is operational, not existential: a hosted builder is going away, OpenAI has named replacements, and the dates are months out. The only mistake is doing nothing until late November.

Frequently asked questions

Sources

Primary — OpenAI

• developers.openai.com — Deprecations — the Agent Builder and Evals deprecation entries, the Nov 30, 2026 shutdown date, the Oct 31, 2026 Evals read-only date, and the “ChatKit remains available” line.
• community.openai.com — Deprecation notice: Agent Builder — developer-forum thread carrying the notice and the ChatKit Advanced-integration migration path.
• developers.openai.com — Agents SDK guide — the recommended code-first migration target for Agent Builder workflows.
• openai.com — Introducing AgentKit — the original DevDay launch (Oct 6, 2025) describing Agent Builder, ChatKit, Connector Registry, and Evals.

Corroborating coverage

• TechCrunch — OpenAI launches AgentKit — independent account of the launch and its components.
• Johann Rehberger on X — early practitioner note on the deprecation and the two migration paths.
• promptfoo.dev — the open-source evaluation framework OpenAI’s Evals migration guidance points to.

Internal links

• /blog/what-is-mcp — the open standard your tool layer should sit on.
• /blog/langgraph-vs-crewai-vs-letta-vs-autogen-2026 — framework-agnostic orchestration options.
• /blog/cowork-vs-chatgpt-agents-vs-operator-vs-manus-2026 — the consumer agent products compared.
• /blog/openai-secure-mcp-tunnels-guide-2026 — securing MCP connections.
• /servers — browse all MCP servers