OpenAI is rolling out several upgrades for developers building agents with its Agents SDK.
The SDK is now available for TypeScript, matching the features of the Python version, including fine-grained control mechanisms and options for human approval. A new RealtimeAgent feature lets developers create voice agents that can run directly in the client or on the server.
These agents now support voice interruptions and can call tools as part of a conversation. The Traces dashboard has also been updated to show real-time data from voice agents, including audio streams, tool usage, and interruptions.
OpenAI says it has also improved the underlying language model. The updated version follows instructions more accurately, calls tools more reliably, and handles interruptions more gracefully. The new model is available as "gpt-4o-realtime-preview-2025-06-03" through the real-time and chat API.
OpenAI's Agent SDK is designed to make it easier for developers to build, deploy, and manage AI agents by simplifying multi-agent workflows, offering configurable language models, agent handoffs, built-in safety controls, and analysis tools. The SDK works with both OpenAI's APIs and other providers' models that follow the Chat Completions standard.
OpenAI also plans to phase out the existing Assistants API in favor of the new Responses API by mid-2026, once feature parity is achieved. The Chat Completions API will continue to receive updates and support for developers who don't need integrated tools, but OpenAI recommends transitioning to the Responses API for expanded capabilities.
