Summary
Charlie Guo from OpenAI’s developer experience team makes a counterintuitive argument: voice agents don’t have to talk back. In a talk aimed at developers building on OpenAI’s audio models, he introduces a three-mode taxonomy — speech-to-speech, speech-to-action, and event-to-speech — and argues that the most powerful and underexplored territory lies in the latter two categories.
Speech-to-action, where voice input drives tool calls and application state changes without generating spoken output, is presented as a vast capability overhang. Guo argues that decades of existing software — APIs, React hooks, database calls — can be trivially wired as tools for a voice model, enabling users to drive entire products with spoken commands. Rather than requiring the model to verbally confirm every action, developers can use existing UI affordances: notifications, visual state changes, cursor animations. The talk draws on examples including government form-filling by voice and creative tools that respond to spoken direction.
Event-to-speech inverts the model: the system speaks proactively when something demands the user’s attention. Guo frames this as an accessibility-first design pattern and a natural escalation layer above notifications and popups. Throughout, he emphasizes that these three modes are remixable, not mutually exclusive, and that the best voice products combine all three. The talk is notable for coming directly from someone inside OpenAI whose job is watching how developers actually build with these models.
📺 Source: AI Engineer · Published September 15, 2026
🏷️ Format: Deep Dive
