October 2026 Summaries

Personal AI agents are expected to become a common way for customers to interact with businesses, handling tasks such as scheduling appointments, managing billing issues, and making routine changes without requiring customers to join calls. Vapi’s beta Personal Agent Calling service aims to support this model, emphasizing that businesses must be able to verify whom an agent represents, what authority it has, and whether its permission remains valid while avoiding unnecessary handoffs to the customer. The proposed approach distinguishes between actions already authorized, such as moving an appointment within specified limits, and requests requiring renewed approval, such as authorizing repairs or accepting an unsuitable time. Emerging standards including Agent2Agent and OAuth could help agents, businesses, and internal systems exchange structured requests, permissions, and confirmations, while maintaining auditability and clear outcome reporting for both customers and companies. Vapi argues that businesses should begin by supporting a common delegated request, defining acceptable proof of permission, and using real-world calls to identify which protocols and capabilities are needed next.

Vapi has expanded its GPT-Live beta, a full-duplex speech-to-speech model from OpenAI, to all customers after an initial limited testing period with selected teams. Unlike conventional voice-agent pipelines that separately use speech recognition, language models, and text-to-speech, GPT-Live directly processes and produces audio while maintaining continuous listening, enabling more natural pauses, interruptions, backchannel responses, tone recognition, and lower perceived latency. On Vapi, GPT-Live operates through a speaker model that manages the conversation and a separate reasoner model that performs procedures, tool calls, and lookups when prompted to do so. Customers can migrate existing assistants by separating conversational guidance into a speaker prompt and operational procedures into a reasoner prompt, while retaining existing phone numbers, tools, testing, recordings, transcripts, and observability features. Vapi positions speech-to-speech models as an additional option alongside modular voice pipelines, which remain useful for teams needing independent control over transcription, intelligence, and voice components.