OpenAI launched AgentKit, a full-stack platform for building production-grade AI agents, featuring the Agent SDK, visual Builder, eval frameworks, and human-in-the-loop approvals. Christina Cai demonstrated creating a customer support agent in under 8 minutes using the no-code Builder, while Sherwin Wu explained the strategic adoption of Anthropic’s MCP protocol for universal tool connectivity. The team emphasized visual workflows as essential for managing complex, multi-step agents and introduced automated prompt optimization—like zero-gradient fine-tuning—to ease transitions across model updates. AgentKit is already powering OpenAI’s own help desk, showcasing internal dogfooding. New tools like ChatKit and the Service Health Dashboard aim to standardize interfaces and ensure five-nines reliability. The evolution from plugins to GPTs to full agent platforms reflects a deeper commitment to scalable, reusable components, with Codex accelerating internal development and shared UI modules reducing redundancy.