LISTENING · SPEAKING · SHARED CONTEXT
Full-duplex AI, put to work.
A conversation that can continue while the work happens.
Full-duplex voice AI can receive speech while producing speech. Listening and speaking can overlap, giving people room to interrupt, clarify or respond during a conversation. Full-Duplex Code applies that interaction to working with a coding agent.
Listening and speaking can overlap
A turn-based voice interface typically waits for one side to finish before the other replies. In a full-duplex conversation, both audio directions remain available together. A person can begin a correction during an explanation, or respond as a thought develops.
Overlapping speech is a capability, not a measure of how well an agent understands you. The experience also depends on timing, useful responses, audio quality and whether the agent keeps track of the task.
Natural turn-taking is part of the interface
Conversation includes pauses, acknowledgements and changes of direction. For a coding task, those moments can carry practical information: “yes, continue”, “wait, use this example” or “show me what you mean”.
Full-Duplex Code gives that exchange a place alongside the coding process. You can stay engaged with the live companion while implementation runs, then respond to the browser result when it appears.
Speech-to-speech, full-duplex and coding agents
These terms describe different parts of a system. Speech-to-speech describes an audio input and audio output interface. Full-duplex describes when those audio directions can operate. A coding agent uses tools to inspect or change software.
Full-Duplex Code is an agent harness connecting a live voice companion, a coding harness and presentation tools. The project is about how people work with an agent through conversation and visuals.
| Term | What it describes |
|---|---|
| Voice dictation | Spoken words become written input. |
| Turn-based voice interaction | One side finishes before the next reply. |
| Full-duplex conversation | Listening and speaking can overlap. |
| Multimodal coding interaction | Conversation is paired with browser visuals and coding tools. |
A shared browser gives the conversation context
An explanation is easier to follow when you can see what it refers to. The agent can show an interface, point to part of a page or present a simple diagram while speaking. You can ask about the thing on screen instead of reconstructing it from a long spoken answer.
Duplex Party streams the browser or terminal to the caller and the audience. The visible transcript retains both sides of the conversation, including interruptions. It is a place to experience voice interaction around real coding work.
Try an interruptible coding conversation
Watch the shared computer, then join the queue if you want to take the microphone. Ask the agent to explain something already on screen, make a small change or show you what previous visitors built.
The live experiment starts with one caller at a time. Projects persist between turns, so conversation can lead to something another person picks up later.
Questions about full-duplex AI
What does full-duplex mean in voice AI?
It means a system can receive and produce speech at the same time. This supports overlapping conversation rather than requiring listening and speaking to alternate.
Is full-duplex the same as speech-to-speech?
They describe different properties. Speech-to-speech concerns audio input and output. Full-duplex concerns whether listening and speaking can overlap.
Is Full-Duplex Code a speech model?
Full-Duplex Code is an agent harness that connects a live voice companion to coding and presentation tools. Duplex Party uses it with Codex CLI and Open Claude in Chrome on a shared computer.
Can the agent show things while talking?
The shared browser is the primary presentation interface. The agent can use Open Claude in Chrome to show pages, diagrams, apps and results alongside its spoken explanation.