Hello,
I am refactoring our company’s voice agent orchestration and am considering adopting Agents SDK, but I’m just afraid it’s not mature enough (especially for voice applications).
I would be happy to learn about your experience.
Thanks a lot!
Hey @f10w, how's it going?
That’s valid, especially for voice. As you're already aware, voice agents are less forgiving than text agents because latency, interruptions calls and recovery behavior all affect the user experience immediately.
I would not dismiss the Agents SDK outright. Also the documentation says the agent SDK is a recommended starting point for voice-agent development, especially with Realtime API. The SDK is designed to help with tools, guardrails, handoffs, session history, and tracing, while still using the Realtime API underneath.
That said, I would recommend to adopt incrementally, not as a full orchestration from the beginning.
Let me know if that helps.
- Avinash
Hi @f10w!
I am probably biased here, but I wanted to add a few details.
The Agents SDK did get a major recent update, centered on Sandbox Agents, controlled execution environments, memory/resume, and harness customization. The caveat is that this is still version 0.14.0.
Just a few days ago, the default model for realtime agents was updated to the newest model, gpt-realtime-2.
In the context of using the newest realtime models, I suggest following the best practices guidance. I recently created a list of relevant sources and will link it below.
Hope this helps!
https://openai.com/index/the-next-evolution-of-the-agents-sdk/