OpenAI AgentKit (openai.com)

OpenAI today launched AgentKit, an integrated toolkit for building, deploying, and optimizing agentic applications—bundling a visual workflow designer, connector governance, embeddable chat UI, safety tooling, and expanded evaluation and fine-tuning capabilities. The suite aims to remove the fragmented, error-prone toolchains developers have used until now by providing an Agent Builder visual canvas with versioning, preview runs and inline evals; a Connector Registry to centralize enterprise data sources (Dropbox, Google Drive, SharePoint, Teams, etc.); and ChatKit for streaming, threaded, brandable chat experiences. Early adopters report dramatically faster iteration and improved outcomes (e.g., Ramp cut cycles by ~70%, Klarna and Clay achieved major automation/growth wins). Technically, AgentKit extends OpenAI’s Responses API and Evals: new eval features include datasets, trace grading for end-to-end workflow assessment, automated prompt optimization, and third‑party model evaluation. Guardrails—an open-source, modular safety layer (Python/JS libraries)—can mask PII, detect jailbreaks, and enforce policies. Reinforcement fine-tuning (RFT) is GA on o4-mini and in private beta for GPT‑5, now with custom tool-calls and custom graders to teach models when to invoke tools and how to be judged. ChatKit and the Evals upgrades are generally available; Agent Builder and Connector Registry are in beta/rolling enterprise rollout. Together these components streamline production-grade agent development, governance, and continuous evaluation across models and orgs.