What we need to make voice AI agentic (www.ultravox.ai)

Recent discussions highlight the urgent need to evolve Voice AI systems from their current, outdated state, which resembles traditional IVR systems, towards truly agentic frameworks by 2026. Despite advancements in text-based models like Claude and GPT-4, voice AI still lags due to significant inference times and outdated training techniques. The most widely used models in production, such as GPT-4o and Gemini 2.5 Flash, exhibit latencies that can hinder real-time communication, making interactions feel robotic and awkward. As younger, more capable models demonstrate slower reasoning, teams are forced to rely on older systems to maintain faster response times, sacrificing intelligence for speed. To develop truly agentic voice AI, three key principles must be met: speed, fluidity, and fluency. A successful system must achieve sub-one-second latency to satisfy user expectations, while also managing complex conversation states without affecting performance. This necessitates models capable of seamlessly integrating tool calls and handling the nuances of human dialogue. The Ultravox initiative aims to address these challenges, presenting fast and intelligent speech-native solutions that prioritize a natural conversational experience. Upcoming articles will delve into the design patterns required to create these advanced voice AI systems.