OpenAI launches AgentKit to build and fine-tune AI agents

Building a reliable AI agent has long felt like assembling IKEA furniture with missing screws and no manual. You had to glue together disparate orchestration frameworks, manually build evaluation scripts, and spend weeks wrestling with frontend components just to get a basic chat box working.

At DevDay on Monday, OpenAI rolled out AgentKit—a consolidated suite of developer and enterprise tools designed to drag agent development out of the hackathon phase and directly into production.

Sam Altman took the stage to frame the launch as an end to developer friction, with live demos and enterprise proof points meant to show that autonomous workflows are ready for the real world.

The Canva for multi-agent workflows

The centerpiece of the toolkit is Agent Builder, currently in beta. Altman described it simply as "Canva for building agents." Instead of wrestling with hundreds of lines of glue code, developers drag and drop nodes across a visual canvas to link multiple agents, define routing logic, and inspect version histories.

To drive the point home, OpenAI engineer Christina Huang built a full workflow and deployed two interacting agents live on stage in under eight minutes.

Early enterprise partners backed up the speed claims. Ramp reported slashing its iteration cycle by 70%, shipping a buyer agent in two sprints rather than two quarters. Japan’s LY Corporation claimed it built and executed a multi-agent workplace assistant in less than two hours.

Alongside the canvas, OpenAI launched ChatKit, an out-of-the-box UI component that is now generally available. It handles messy frontend hurdles like streaming text and displaying model "thinking" traces while letting teams drop a branded chat interface straight into their web apps.

For enterprise oversight, the new Connector Registry acts as a central hub across ChatGPT and API workspaces to manage hooks into Google Drive, SharePoint, Teams, and third-party MCP servers.

Leveling up evals and reasoning fine-tuning

Building an agent is easy; knowing whether it will hallucinate in front of a customer is hard. OpenAI addressed this by expanding its Evals platform, which is now generally available.

The update introduces trace grading to pinpoint exactly which step failed in a multi-turn chain, automated prompt optimization based on human feedback, and a surprising concession: developers can now benchmark third-party models right inside OpenAI’s platform.

For advanced builders, Reinforcement Fine-Tuning is now generally available on o4-mini and in private beta for GPT-5. The beta introduces custom tool calling and bespoke graders, effectively letting teams train models when and how to fire external APIs.

Notably, OpenAI later updated its documentation to state that the visual Agent Builder and standalone Evals will be wound down on November 30, 2026, advising code-first builders to stick with the Agents SDK and prompt-first teams to use Workspace Agents.

Why it matters

We think AgentKit is a clear sign that OpenAI wants to own the entire agent stack, not just sell token compute by the pound.

For developers, ChatKit and built-in tracing eliminate weeks of thankless plumbing. If you are an enterprise trying to deploy internal knowledge bots or automated support, the barrier to testing a working prototype has basically vanished.

For the broader ecosystem, this puts direct pressure on middleware players like LangChain, Flowise, and specialized evaluation startups. When the provider supplying your intelligence also gives you the canvas, the connectors, and the UI for free with standard API pricing, paying for third-party wrappers gets a lot harder to justify.

Yet OpenAI's later decision to sunset Agent Builder in favor of its code-level Agents SDK reveals the underlying catch: low-code drag-and-drop tools are brilliant for eight-minute keynotes and non-technical stakeholders, but serious production engineers almost always retreat to pure code when edge cases break down.