OpenAI has recently updated its AI agent development tools, focusing on expanding platform compatibility and improving voice interface capabilities. These enhancements aim to facilitate the creation of more practical, controllable, and auditable AI agents for real-world applications.
The Agents SDK is now available in TypeScript, extending support to JavaScript and Node.js developers. This aligns the SDK with modern web and cloud-native application stacks, enabling agent deployment across both frontend and backend environments. The TypeScript SDK mirrors the features of the Python version, including support for handoffs, guardrails, tracing, and the Model Context Protocol (MCP).
RealtimeAgents extends the Agents SDK with audio input/output, stateful interactions, and interruption handling specifically designed for voice applications. Human-in-the-loop (HITL) approval allows developers to intercept agent execution for manual confirmation, which is crucial for oversight and compliance. OpenAI has also updated its speech-to-speech model to reduce latency and improve the naturalness of real-time audio interactions. The updated model is available as gpt-4o-realtime-preview-2025-06-03 in the Realtime API and gpt-4o-audio-preview-2025-06-03 in the Chat Completions API.