In October 2025, at DevDay, Sam Altman walked on stage and pitched AgentKit as the thing that would take agents "from prototype to production." Its centerpiece was Agent Builder, a visual drag-and-drop canvas for wiring up multi-agent workflows — Altman's "Canva for building agents." Around it sat ChatKit (an embeddable chat UI), a Connector Registry, Guardrails, and a beefed-up Evals platform.

On June 3, 2026, OpenAI posted a deprecation notice for Agent Builder. It shuts down November 30, 2026. The migration advice: move to the Agents SDK if you write code, or to ChatGPT's Workspace Agents if you don't.

Eight months. That is the whole arc, launch to sunset, for the most-demoed piece of the most-hyped agent release of the year. And the interesting thing is not that OpenAI changed its mind. It's which layer it killed and which one it kept.

The part that died and the part that lived

OpenAI did not deprecate AgentKit wholesale. ChatKit is still around. The Guardrails and Evals work continues. What got the notice was specifically the visual authoring surface — the canvas where you assembled an agent by dragging boxes instead of writing code.

The drag-and-drop layer was the demo. The code layer was the product. Only one of them survived its first year.

That split is the entire lesson. A visual builder is, by construction, an interpreter for a graph that lives in the vendor's runtime. You don't own the artifact; you own a configuration of someone else's hosted execution engine. When that engine's roadmap shifts, your workflow shifts with it — or disappears. Agent Builder could export to Agents SDK code, which softened the landing, but OpenAI itself is careful to call the export a starting point, not a guarantee of identical behavior. A migration, not a download.

What LangGraph is doing differently

LangGraph starts from the opposite premise. The agent is a graph you define in code — nodes, edges, cycles, conditional branches — and that code is the artifact. You run it on your own infrastructure or any cloud. It is model-agnostic: the same graph can call OpenAI today, Claude tomorrow, or a self-hosted open-weight model when finance asks you to cut the inference bill. Nobody can deprecate your while loop.

This is not a knock on visual tools in the abstract. Agent Builder's node vocabulary was genuinely thoughtful — If/Else for branching, While for looping until a condition, a User Approval node for human-in-the-loop, Set State for globals, and first-class MCP nodes so external tools came in over a standard protocol instead of bespoke webhooks. For a prototype, dragging those boxes is faster than scaffolding a graph by hand.

The catch is that the rigidity that makes a canvas legible also makes it shallow. Branching is manual; the model isn't deciding the control flow so much as falling through hand-placed gates. The moment your agent needs to do something the node palette doesn't express, you're exporting to code anyway — which is to say, you're back in Agents-SDK or LangGraph territory, just later and with a translation step in between.

The portability hides one level down

Here is the part most "AgentKit vs LangGraph" comparisons miss. The honest comparison isn't AgentKit-the-canvas against LangGraph. It's OpenAI's code-first Agents SDK against LangGraph — and at that layer, OpenAI is more open than its branding suggests. The Agents SDK uses the Responses API by default for OpenAI models, but it is provider-agnostic: it can drive Anthropic, Gemini, Mistral, and 100+ others through LiteLLM. The lock-in was never really in the SDK. It was in the hosted visual runtime sitting on top of it — exactly the layer that just got a shutdown date.

So the deprecation isn't a story about OpenAI failing. It's a clean natural experiment in where durability lives. The closer a layer sits to "configuration of a vendor's hosted engine," the shorter its expected life. The closer it sits to "code I run wherever I want," the longer. That ordering held even inside a single company's own product line, in under a year.

How to choose now

The decision is less about features than about time horizon.

The broader pattern is worth holding onto, because it keeps recurring. OpenAI's bet with AgentKit was to own how you author an agent. AWS, with Bedrock AgentCore, made the opposite bet: stay neutral on the framework and the model, and own where the agent runs. One of those bets just had its centerpiece deprecated eight months in. The other is selling the thing that doesn't expire when the roadmap turns — the substrate. When you choose a layer to build on, choose the one whose incentives line up with still existing next year.